# 影片筆記：The ULTIMATE Beginner’s Guide to OpenAI Codex (3-Hour Course)

## 處理狀態

已完成語音辨識與逐字時間軸。自動摘要階段未在時間預算內完成，因此先發布可用版本；可於後續從現有逐字稿補做摘要，不需重新辨識影片。

## 來源

https://www.youtube.com/watch?v=lmGX79cumyw

## 完整逐字稿

In this video, we're going to do a comprehensive Codex tutorial.
I want to show you not just what Codex is, but how to actually use it as a serious app building agent.
We'll start with the foundations, how Codex works, how the Mac app and the CLI fit together, how to choose models and thinking modes, how threads and contexts work, and how to use tools like the browser, terminal, files, and external systems.
Then, we'll turn that foundation into a real workflow.
We'll use Codex to plan, architect, design, implement, debug, and ultimately deploy a production-ready app from scratch.
By the end, you should understand the full Codex workflow well enough to take your app idea, break it into manageable steps, and use Codex to build it with much more confidence.
Now, when you Google for Codex, you're going to land on this website here.
And here, you can learn more about Codex, and you can also download the app.
Now, as you can see, it says here download for macOS.
That is because I have macOS.
If you're running Windows, you're going to have a button that says download for Windows.
So, it does support macOS and Windows.
And when you scroll down, you can see the main screenshot that kind of gives you a general overview, general idea of what this app looks like, okay?
And you can kind of scroll down, you can learn more about it.
Or we can jump into my Codex instance here, and I can give you a quick overview.
So, on the left-hand side, we have things like starting a new chat, search, various plugins, automations.
Just below that, we have our main organization, right?
And this is where you can organize your projects, okay?
And the idea here is that anytime you have a project, you want to click here, and you want to either start it from scratch or you use an existing folder because these are not organized just logically.
They're actually linked to folders on your disk, okay?
And so, everything that you see here represents actual folders.
Now, inside of these folders, we have some items here.
Now, these are not files, and they do not represent anything on the disk.
These are actually conversations, right?
And you can think of them as, you know, similar to what you might have in Chat GPT, okay?
So, if you're working on this project, you can create a new conversation simply by, you know, clicking here or pressing command and and telling it what you want to happen.
Now, in the center pane over here, this is where you're going to be communicating, right?
So, if we scroll back up, this was my initial prompt here.
I want to build a very, very simple web app that arranges items as a Trello Kanban board style, etc.
And then I have this interaction here, right?
It does something, I ask it to do something else, etc., etc.
Now, just below that is where you're going to be entering your prompt, okay?
So, you're going to be selecting the model, uh the different reasoning, and we're going to be covering all that, but this is the main text box for the model.
Now, on the right-hand side, there's a lot of interesting things happening here that a lot of people are not really aware of, okay?
And a lot of people are not really like using it.
And And you need to understand to leverage this whole app experience, right?
So, right here, we have this little card that displays a little bit of information about this specific conversation, right?
So, we have branch details, we have some artifacts, right?
And these are the data that happened as a result of this convo, right?
And so, if I go to another conversation, we might have different artifacts, right?
And you can think of an artifact as a side effect of, you know, of your conversation, okay?
So, if you've created a readme file, you're going to see it here.
If you did a web search, you're going to see it here.
Now, one of the more interesting things happens when you click this button and open the right-hand pane.
And here, you can see all kinds of interesting things, right?
So, for instance, if I click this button, I can have a side chat, or I can have a browser.
Now, the browser feature is actually very, very powerful, because I can simply open this up here, and as I'm developing this app, it's going to give me a URL that I can open here in this browser, and that will allow me quick iteration, because I'll be able to
see exactly how the app looks as I'm developing it.
And we're going to be doing all of that in more detail a little bit later in this tutorial.
First, let's talk about Codex.
The easiest way to understand Codex is that it starts from the same basic idea as ChatGPT, but it is connected to your actual working environment.
So, instead of only answering questions, it can inspect files, edit code, run commands, use the browser, test the app, and help move a project forward.
The difference matters a lot.
ChatGPT is great when I want an explanation, a rewrite, a brainstorm, or a quick answer.
Codex is what I use when I want work to happen inside a real project.
It can do many of the same reasoning tasks as ChatGPT, but it can also act on the project directly.
That is why I keep ChatGPT mostly for pure informational questions, and I keep Coda for work threads.
So, my project context stays focused and clean.
There are two main ways to use Coda.
The first is through the CLI, which is best when you already live in the terminal and want to run quick repo tasks, scripts, lightweight edits, or developer style iterations.
The second is the Mac app, which is usually better for visual app building work.
If I want to see the browser preview, inspect file changes, manage a longer project, approve permissions, review artifacts, or teach the workflow on screen, the Mac app gives me a much clearer workspace.
And that is why I highly prefer to use the Mac app, and that is what I recommend you use as well.
Inside the Mac app, the main pieces to understand are the project or the workspace, the thread area, the model selector, the thinking mode selector, file changes, terminal output, browser preview, artifacts, permissions, and the review flow.
Once those pieces make sense, Coda stops feeling like a chat box and starts feeling like an app building environment.
The project pane is where you open or switch between projects.
The thread area is where the conversation and task history lives.
The file change view shows what Coda wants to edit before those changes are accepted.
The terminal output shows commands, builds, errors, and verification steps.
The browser preview is where you can inspect the running app is where you can inspect the running app visually.
Together, these pieces make Coda feel less like a chat bot and more like a working development environment.
Permission flow is also important.
Cortex should never feel like a black box that silently changes everything.
As soon as it finishes running commands, editing files, or taking an action that matters, you can always inspect what it's doing, approve the right actions, and keep control over the project.
The next thing to understand is that Cortex is not just one fixed experience.
There are a few workflow decisions you make each time you use it, and the first one is which model to use.
In practice, I usually default to the latest strongest model unless I have a clear reason not to.
Because most app building tasks benefit from better reasoning.
Here's a rough guide to picking the correct model for your specific task.
So, for architecture, back-end design, multi-step implementation, difficult debugging, or final review passes, I want a stronger model, such as GPT-5.5.
For small copy tweaks, simple UI polish, short refactors, or low-risk tasks, a faster, a lighter model can be a better tool.
And here's where you can use an older model, or a mini, or a nano version of the current latest model.
For example, if I'm asking Cortex to rename a button, tighten some copy, or adjust spacing on one component, I do not need to overthink the model choice.
But if I'm asking Cortex to design the Convex schema, think through permissions, restructure several files, or debug a problem that spans front-end and back-end code, That is where I want the strongest model That is where I want the strongest model available.
The more expensive the mistake, the more I want to slow down and use the better reasoning setup.
The next important decision is the reasoning mode.
This controls how much time and effort Cortex spends thinking through the task before and during the work.
This decision is a little more nuanced than model choice because every reasoning mode has a place.
Low thinking is useful for quick, obvious tasks where the risk is small.
Medium is good for normal app building work, first pass UI normal app building work, first pass UI implementation, straightforward debugging, and routine feature work.
High is what I use when the decision is more expensive.
Architecture, convex data model design, authentication, permissions, hard bugs, multi-file changes, and the final review.
My default for serious app building work is usually a strong model with high thinking because that gives me the best balance between speed and the best balance between speed and quality.
But, the practical lesson is simple.
Match the model and thinking mode to the risk of the decision.
If the task is small and reversible, keep it fast.
If the task affects the structure of the app, slow the model down and let it reason.
Now, as you're working on your project and you're in this conversation project and you're in this conversation view, you can enter your prompts here, and just below, you can select various things.
Now, first of all, you have your contacts window here, right?
Here you can see how much of the context you have used.
Now, Korax says that it automatically compacts its contacts.
And yeah, you can do that, but over time, it's not going to be as effective.
And so, you always want to be starting new conversations anytime you're working on something new that requires a context switch, okay?
Next, we have here the all-important model selector, okay?
So, if I click over here, I have it automatically set to GPT-5.5, which is the newest model as of the time of recording this video, okay?
You can also switch it to some of the older models as well, but I see no reason to do that because this model has been very, very capable.
We also have things like speed, standard versus fast.
Fast is 1 and 1/2 times faster, but it also uses more of your usage, okay?
So, depending on which plan you're at, you may want to stay on standard.
I mostly keep it on standard unless, you know, I'm in a hurry and I want to switch it on fast.
Next, we have intelligence, right?
Now, this is very, very important.
By default, I keep this on high.
I found that to be a good balance between speed and quality because as you go towards lower intelligence, the model is going to be faster, but you're going to be giving up quality, right?
You're going to be giving up intelligence.
On the other hand, as you go into higher and higher intelligence, the model is going to be slower, but it's going to be thinking more, right?
It's kind of like a person, right?
You can ask a person a question, and that person may think more, think through something longer, right?
You know how people say, "Let me sleep or let me sleep on it." Or let me think it over, right?
The longer you think, the more the more you figure something out, and the the the the the higher chance you're going to arrive at the correct answer, at least for you.
You're going to think through all the possibilities, whereas the first answer might be wrong, and usually is because you haven't thought everything through.
And it's the exact same thing here, right?
If you set something too low, and you're working through something fairly complex, well, the model is not going to think it through.
It's going to give you an answer that well, it's an answer, but it might not be the ideal answer, okay?
On the other hand, with extra high, you know, the model is going to think longer, but it may not be necessary in a lot of situations.
And so, I typically keep it on high, sometimes medium, very rarely low, and also very rarely extra high, unless I'm working through something that the model has trouble with high.
You know, the situations where extra high really proven itself over high, it hasn't been that many.
And so, for the most part, I keep it on high, sometimes I also switch it back to medium.
And as you start to use Codex more and more, and as you begin building more and more apps with it, and you'll be able to better judge which thinking mode to better judge which thinking mode to select.
Now, before we go deeper into the workflow, I want to quickly talk about pricing and plans because this is one of the first practical questions people have with Codex.
For most people learning Codex, the important question is not what is the most expensive plan.
The question is, what do I actually need for the kind of work I'm doing?
If you're learning, testing prompts, and building smaller projects, you can usually start with the normal Chat GPT plan access available to you.
If you're doing heavier app builds, longer agent runs, team workflows, or professional coding work, then plan limits, credits, and business options start to matter more.
Now, here I am on OpenAI's Codex pricing page.
And if you scroll down, you can see the prices that are available at the time of filming this video, right?
So, we have the free plan, and this allows you to explore Codex capabilities on quick coding tasks.
We have the gold plan, which is $8 a month, and you can use Codex for lightweight coding tasks.
We have the plus plan, which is fairly popular, 20 bucks a month, and here you can finally start to actually do something meaningful, right?
Something like, you know, building a part of an app, or building some scripts.
And they say this can power a few focused coding sessions each week.
Next, we have the pro plan, right?
And so, here you can choose whether you want the 5x or the 20x higher rate limits than plus, and this is $100 a month.
I'm actually using the pro plan 20x.
Okay, so I'm paying 200 bucks a month, and at this level, I am at a point where I pretty much never reach any limits.
And that means I can build, you know, multiple projects at once, and use Codex for all kinds of tasks, and still not reach limits.
And so, during the time where I'm when I'm, you know, building heavy apps and and doing some serious work, I might get close, right?
But I never really get to the limit.
You need to decide which plan is right for you.
Now, from what I understand, plus plan is very, very popular.
And so, what I would do is I would obviously, you know, try the free plan, try to see how that goes, and, you know, obviously, you know, if you do anything serious or remotely serious, you're probably, you know, you're going to run out run out of
this plan.
And so, the next logical step is to try out the plus plan.
The plus plan is where a lot of people start and it gives you a reference point as to, you know, the amount of work that you're doing, how much can this plan actually cover the amount of work you're doing.
Because the fact of the matter is there people that work on very heavy projects and, you know, there they work on big repos, right?
Lots and lots of code base.
They run out of their plan very, very quickly.
But if you're working on a smaller projects, you're not using the fast mode.
So for instance, you're not coming here and you're changing you're not changing speed because I'm using the fast mode from time to time and that, you know, increases the usage as well.
So if you're on standard and you're using maybe medium often, right?
You're not really using high that much and you're almost never using extra high, you can get a lot of mileage out of this plus plan.
And when this plus plan starts to run out, you know, you're you're starting to notice some tight limits, you may want to switch to a higher plan, okay?
Such as the pro 5x or 20x.
Now you can also ask Codex to help you choose the right Codex plan for your specific use case.
For instance, you can use this prompt here, help me choose the right Codex plan for my use case, okay?
Use only current official OpenAI pricing and help center information.
And I'm going to go ahead and run this and see what it says.
All right, then this is what we got back.
As of this day, ChatGPT plan is that include Codex, okay?
So this is the practical read.
Free, try Codex on quick tasks.
Go lightweight coding tasks.
Plus a few focus coding sessions each week, okay?
And this is exactly what I just told you from that official page.
And so if you scroll down, you can see how Codex usage and credits work.
You can learn you can read this.
Best fit for your use case, okay?
For learning Codex and building small to medium apps, ChatGPT plus is the sensible default.
If your current plan is Plus, I would not upgrade until you actually hit Codex limits often enough that the friction matters.
Okay, when heavier or business plans make sense, what to verify live before deciding?
And so, a sensible thing to do would be to try Codex Plus and see if that's enough for your specific use case.
If you're not a heavy Codex user, if you just want to experiment, it should be more than enough.
And then you can decide whether you want to upgrade to the higher tier plans.
Now, before we move into the app building workflow, I want to take a moment to talk about some of the things Codex does that can greatly increase your productivity.
These are features I use all the time across different projects.
Skills, plugins, automations, and MCP along with its tools and connectors.
The best way to think about all that is skills tell Codex how to do a workflow, plugins package capabilities, automations let work happen later, and MCP tools and connectors give Codex controlled access to outside systems.
Once those pieces make sense, Codex feels less like a code editor and more like a project agent.
First, let's talk about one of my favorite features that I use pretty much all the time, skills.
Skills are reusable instructions for how Codex should do a workflow.
They're not external tools by themselves.
They're more like local playbooks that tell Codex the style, process, output format, and validation steps for a specific kind of work.
For example, this course uses skills for explainers, Notion editing, video timelines, and other repeatable production tasks.
Instead of explaining the same rules every time, I can point Codex to the skill and it knows the house style, the workflow, and the mistakes to avoid.
The rule is simple.
Anything that gets used over and over again becomes a skill.
That matters because a lot of real work is not just about getting an answer.
It's about getting an answer in the right format, following the right process, and producing something consistent with the rest of the project.
All right, so here I am in my Codex instance here, and I want to show you some things that you can do with skills.
So, I have here a conversation for an app that I was building.
And if I press the the forward slash, you can see that I have many options here, right?
I can do many, many things.
And as you can see for skills, I have lots and lots of skills.
I have tons and tons of skills over there.
And skills allow me to package a repeatable process.
You can think of it as a function that does something.
It gets some input and it produces an output.
And the goal is to produce consistent output as you've done before.
And so, as an example, let's say if I'm working in this convo here, right?
And I asked the model to do something for me.
So, here I have I want to build a very simple web app that arranges items as a Trello Kanban board style, okay?
And it comes up with this.
Now, right here can package that as a skill.
Now, what do I want to package as a skill?
Well, I want to package, for instance, maybe creating the architecture.
So, I can say something like package the architecture step as a skill, and then I just give it a name of the skill.
Sample arch stack.
Okay?
And that's going to create a repeatable skill that will allow me to generate an architecture for any kind of project that I may be working in the future.
So, I'm going to go ahead and jump into the plan mode in case the model wants to confirm something or clarify something.
And I'm going to go ahead and execute this.
And let's see what happens.
Now, it came back to me and it wants me to answer a question.
Should the sample arch stack package this as a general architecture planning skill or specifically as the Kanban Trello app architecture, right?
So, we want it to be a general app architecture.
So, I'm going to go ahead and save this.
And now it came back with a plan, okay?
Create the reusable Codex skill named sample arch stack under wherever it's storing those skills.
The skill will guide Codex through the architecture step for simple app ideas.
And it's going to package that step into a reusable process.
All right.
And now it tells me that created the skill app.
It contains skill.md and then It contains skill.md and then agents.for.openai.yaml.
Validation passed, etc. etc.
And now what I can do is I can enter a prompt and I can say use this skill for a brand new project management app and it should use this skill to create an architecture stack for this kind of app.
So, it wants to know for the brand new project management app, what should the version one data model support?
We're just going to do local only.
And here it came back with a plan.
Brand new local only project management app architecture, right?
That's the purpose of the skill, okay?
We're going to expand this and it's coming up with this architecture, right?
App shape, data state model, component architecture, test plan, etc.
etc.
And we can say yes and it's going to go ahead and implement all that.
And so, this is actually a fairly valuable skill because that is exactly what you want to be doing in those initial stages when you're building an app, as you're going to see a little bit later.
But, by having the skill, we do not need to explain all of that over and over again.
So, as you're building more and more apps, you can simply reuse this skill from now on.
Now, before moving further, I also want to explain a few terms that can sound similar at first, plugins, automations, and MCP.
They all expand what Codex can do, but they are not the same thing.
Plugins are packaged capabilities or workflows that make Codex better at a specific kind of make Codex better at a specific kind of work.
For example, a Notion plugin can help Codex read and update Notion pages.
A Vercel plugin can help with deploys and project settings.
A Hyperframes plugin can help create explainers, and browser or document tools can help Codex inspect apps or work with files.
The simple way to think about plugins is this.
They give Codex a more specialized tool set for a specific job.
Codex still needs the right context and the right instruction, but the plugin gives it a cleaner way to operate inside that workflow.
Automations are different.
A plugin gives Codex a different.
A plugin gives Codex a capability, but an automation gives Codex a time-based job or follow-up behavior.
That might be a reminder, a recurring check, a scheduled review, or a request to come back to a thread later and continue the work.
This is useful because not every task needs to happen immediately.
Sometimes, you might want Codex to check something again tomorrow, remind you to review an outline, monitor whether a deployment is still working, or wake this thread back up when you need it to continue.
Automations turn Codex from a one-time assistant into something that can help manage longer-running work.
The important distinction is simple.
Plugins are about what Codex can do, and automations are about when Codex should do it.
All right, so here I am back in Codex, and I want to show you how plugins work, right?
So, if you click over here on plugins, you can see all the plugins that you have here, okay?
And so, it it tells you that make Codex work your way, and if you scroll down, you can see all the different plugins, right?
So, you can click on the plugin, and you can learn more about it, right?
So, for instance, this spreadsheets allows you to create and edit spreadsheet files, right?
You can use this plugin to create a spreadsheet to analyze this company's financial, etc., etc., right?
Create, edit, inspect, render, verify, and export Excel SX, CSV, and TSV spreadsheets locally.
Use spreadsheets when the durable output or target is a spreadsheet workbook, etc., etc.
So, this includes the skill, right?
So, you can think of package as kind of an umbrella that includes a skill, but it also gives Codex an ability to to connect to external systems, right?
It's kind of this higher-level structure, okay?
Whereas a skill is a very, very specific thing.
It's a specific piece of instruction.
Essentially, a specific prompt.
A plugin is this umbrella that contains a skill, right?
So, you can have that.
One plugin that I use all the time is Notion.
So, if we click over here, you can see that this actually includes a lot more interesting things, right?
Notion workflows for specs, research, meetings, and knowledge capture, right?
So, you can specify this, and then you can give it a prompt.
And if you scroll down, you can see that it includes the Notion app, the knowledge capture.
These are skills.
This is an app.
Okay, so this allows it to connect to an external system, right?
So, if I click over here, I can manage this in ChatGPT as well, because it's an external app.
And this right here, these are skills, right?
I can open that, and that's going to open the skill right here, right?
This is This is how the skill is actually stored.
And I can also turn it off or turn it on, depending on if I want that skill enabled or not.
Now, this is a plugin that I use all the time, right?
I use Notion for my planning, for, you know, writing stuff, scheduling stuff like that.
It's the brain of my workflow, essentially.
And then I installed this pre-built Notion plugin, and the first time I ran it, it authenticated against my Notion workspace, right?
Connected to my Notion account.
And from then on, I can simply say, "Hey, you know, use this Notion spec to do this, or to do that, or, you know, edit this document, edit that document." And it's able to go out and edit the right documents, and essentially make Notion an extension of
Codex.
That's kind of the purpose of a skill, to make something an extension of Codex.
Right?
And that's the main idea with all of this, right?
Another really cool one is, like, for instance, Google Calendar is another one, right?
Instead of connecting to Notion, or in addition to connecting to Notion, you can also connect to Google Calendar.
I personally do not use Google Calendar, I use Notion, so I do not have this skill set up.
Another really useful skill that you're definitely going to be using if you're building apps is Vercel, right?
Vercel, or maybe something like Netlify for deploying your apps, right?
I use Vercel when demoing apps.
And by having this package installed, it can figure out how to deploy, how to check deployment, how to do everything without me needing to actually understand how all of that works, without me needing to go to to Vercel and read their docs and
understand how to deploy and all of that.
I simply install this skill and then I you know, I click here and now I can say, you know, Vercel, order this repo for Vercel deployment risk.
Or Vercel, which Vercel tools fit fit this app app best.
Or Vercel, deploy this app, right?
And if you scroll down, you can see that this has this app and then it has all of these skills.
All of these different skills.
Look at this.
It has a ton a ton of skills.
Vercel CLI, Vercel agent, all of these.
And obviously, these skills are not just for deploying, right?
Because we also have payments.
We have best practices.
We have all of these other ones.
So, these skills essentially map to different parts of Vercel, to different features of Vercel.
We're not just talking about, you know, deploying your app, because this could be just a skill, deployment, right?
This could be one skill that I may need.
I I may not need all of these other skills.
But why not?
You know, Codex will figure out which skill it needs to consult, you know, according to the prompt that I give it.
So, it doesn't you know, it's not a big deal to have all of these enabled if I'm using Vercel.
And like I said, this allows me to basically bring Vercel into my Codex workflow.
That's the beauty of these plugins, okay?
And that's the same idea with all of these plugins.
You're essentially bringing an external system and letting Codex manage it for you.
Now, automations work very, very differently.
So, if you click on automations here, you can see that automations, they allow you to automate work by setting up scheduled chats, okay?
And you can learn more here.
Now, if you click on any of these automations, you can see exactly how they're going to be working.
So, if I click on something like summarize yesterday's Git activity for standup, you can see that it wants to create an automation for me, okay?
And what exactly is it trying to do?
Well, it's essentially trying to create a essentially trying to create a conversation that will be repeatable, okay?
So, it creates this automation.
This is the name of the automation, right?
This is the prompt.
This right here is the prompt.
And the prompt is summarize yesterday's Git activity for standup, grounding rules, etc., etc.
We can select a work tree.
We can select a project here.
And then we can select the time where it's going to run over and over again, right?
So, with this kind of automation, you may want to run it on weekdays at 9:00 a.m.
But, you can also select it, make it hourly, daily, weekly, etc., etc.
And you can also change the time, obviously.
Now, here you can change the model, and here you can change the reasoning level.
Now, with this kind of automation, you don't really need a high reasoning.
Medium is going to be just fine.
And after you're done, you can simply click create, and now you're going to have an automation that will run every weekday at 9:00 a.m.
running this prompt.
And And remember, this is a special prompt.
Here, you're asking it to consult Git activity.
And so, the idea is you may have, you know, Git connected via a plugin or something else, and it's going to connect to your Git environment, and then figure something out for you.
And that's going to run every weekday at 9:00 a.m., right?
So, instead of you needing to, you know, jump into Codox first thing weekday morning and execute this prompt manually, you can create an automation to do just that.
If you scroll down, you can see some other automations here.
You have incidents, you have code quality, right?
Maybe you want to run a code quality check every 24 hours.
Well, you can click here daily at 9:00 a.m., right?
Scan recent commits since the last run or last 24 hours for likely bugs and propose minimal fixes.
Okay, that's another very, very useful automation, right?
Repo maintenance.
You have here growth and exploration, right?
Click here from recent PRs and reviews, suggest next skills to deepen.
This is great, right?
So, automations, they could be used for pretty much anything, right?
And they are, you know, it doesn't really matter what kind of niche you're in.
You don't need to be in software development or project management or anything like that.
You can be a writer, right?
You want a status.
You're writing maybe 1,000 words every day.
You know, you can have an automation that checks for spelling, grammar errors that you may have done, and then it may suggest some fixes and also another automation that gives you more ideas, you know, to break through writer's block.
And that can run every morning at 9:00 or 10:00 a.m., okay?
And so, automations are very, very useful if you need some kind of a repeatable structure in your workflows.
And I want to talk a little bit about MCP.
So, MCP or model context protocol is very, very important, okay?
So, if you press command comma or you go into settings over here, you're going to be on this screen.
And if you choose MCP servers, you're going to see all your MCP servers that you have here configured, and you can also see whether they're enabled or not.
Now, an MCP server is basically a server.
It's a remote system that your Codex instance can connect to.
So, as you can see, I have a bunch of MCP servers here that do all kinds of interesting things.
For instance, I have here a Stitch MCP server configured.
I have here a Superbase MCP server.
I also have, you know, Flow Upload Designer, Flow Base, a bunch of apps that support MCP.
And when you connect to them, Codex is acting like an MCP client.
So, you have that same client-server paradigm.
Now, why would you want to use an MCP server to connect to something like Superbase or something else?
Well, the reason is very simple.
You want Codex to do all the administration, all the work for you, right?
So, if you're building an app using something like Superbase, and we're going to touch upon that a little bit later in the video, you can set up Superbase yourself.
You can configure all of that.
That takes time, but it's, you know, it's very, very doable.
Or, you can simply configure a Superbase MCP server in your Codex instance.
And that will allow Codex to use that MCP server to essentially manage this remote server, this tool, this app for you.
And as you're going to see a little bit later in this tutorial, this is very, very helpful to us because it saves a lot of time and it enables things that we couldn't really do before because now Codex understands One Piece and via MCP
it can get more contacts and create a much more robust app for you or a tool for you or anything else, a system, if you will, because now it can connect to other remote systems and essentially consult that data with the data we
already have, such as the code or things it has been working on, and figure out a unified solution.
And this is extremely helpful to us, okay?
And so, there's a couple of ways that you can add a server.
You can click here, and you can fill all of this manually.
Or, you can simply tell Codex, "Set up and activate a Supabase MCP server." It's going to figure out how to do that.
It's going to search for it.
It's going to figure out the specs, and it's going to set it up for you.
And the same thing for all of these other ones.
Now, after setting up the MCP server, you can simply tell Codex to use it.
So, I can say something like, "Use the Supabase MCP server to query data, or check how many tables we have, or pretty much anything else." And it's going to go ahead and connect to your Supabase instance in this example via the configured MCP server.
Now that we have laid the foundation, and before we start building the real app, let's talk about the ideal Codex workflow.
What is the best way to approach it?
Before we begin building our app, you want to establish the workflow that you're going to use again and again.
The loop is simple.
Inspect, plan, edit, verify, and report.
That sounds basic, but it is the difference between using Codex as a random code generator and using Codex as a controlled build partner.
The mistake a lot of people make is that they jump straight to "Build this app." or "Fix this problem." without giving the agent a clean process.
That can work for small tasks, but it breaks down quickly on real projects.
For a serious build, you want Codex to understand the repo first, explain the plan, make a focused change, verify the result, and then report back in a way that you can re-review.
A good Codex prompt is not just a request.
It is a small project brief.
The better the prompt is, the easier it is for Codex to understand the current state, choose the right amount of work, and verify the result.
The basic structure is simple.
Explain the goal, give the relevant context, define what should change, define what should not change, and tell Codex how to verify the work.
If the task is risky or unclear, ask Codex to inspect first and make a plan before editing.
For instance, instead of saying "Build the app," I would say, "Inspect the project first, do not edit yet, identify the existing structure, propose the smallest first version, list the files you expect to touch, and explain
how you will verify." That gives Codex a much cleaner job.
For implementation prompts, I want the request to be specific enough that Codex can act, but not so over-specified that I accidentally force the wrong solution.
I usually describe the outcome I want, the constraints I care about, and the checks that need to pass.
Then Codex can use the actual project context to choose the cleanest path.
As part of this workflow, I want to introduce you to one of the most important Codex features, plan mode.
Plan mode is a superpower that I use for 80 to 90% of my prompts.
There are two main benefits to using the plan mode.
First, it's a form of brainstorming where the model cannot change or alter your files.
Second, it's a way for the model to clarify your requests by asking follow-up questions.
This is especially useful when the request is under specified.
If I say, "Build the booking flow." there are a lot of hidden decisions inside that request.
What fields does the form need?
What status should the booking start with?
Where should the booking appear for the helper?
What should happen after submission?
Plan mode gives Codex a chance to surface those questions before it edits the app.
I also use plan mode as a quality filter.
If the plan is vague, too large, or misses something important, I know the implementation will probably be messy.
If the plan is specific, scoped, and easy to verify, then the actual coding step becomes much safer.
So, you want to use plan mode initially while brainstorming and also as you enhance your app because it gives the model a chance to ask follow-up questions to ensure there's no misunderstanding.
Now, if you're working with an existing repo, the first step should always be to inspect.
Codex should understand the files, the app structure, the current state, and the goal before changing anything.
Then it should plan the work in a way that is small enough to review.
Only then, once Codex is fully in sync with your goals, should it make the actual edits and write code.
Once the code is written, it should verify the code with the browser, terminal, using tests, or the build output.
And finally, it should report what changed and what still needs attention.
Verification is the part I want to emphasize.
A task is not finished just because Codex wrote code.
For UI work, I want to see the browser.
For back-end work, I want to see the function call, the dashboard row, or the test results.
For deployment work, I want the live URL and a real smoke test.
That habit keeps the workflow grounded in evidence instead of trusting that the code is probably fine.
Additionally, you always want to break big tasks into smaller one.
Instead of asking Codex to build the whole marketplace as one prompt, you want to ask it to plan the app, build the first UI shell, wire one initial database backed version, add the booking flow, add the helper dashboard,
and so on.
Each step has a clear target and a clear way to check whether it worked.
A good Codex task should usually fit into one reviewable outcome.
Create the first static marketplace homepage is reviewable.
Wire service listings from Convex is reviewable.
Add booking creation and verify the booking row appears is reviewable.
Build the whole app is not reviewable because there are too many decisions happening at once.
So, the workflow is not just a technical process.
It is also a way to stay in control.
We let Codex do meaningful work, but we keep the work divided into chunks where we can understand the change, approve the direction, and catch problems before they spread.
And so, here I am back in Codex, and you can enable plan mode by pressing shift tab, and here you have plan mode.
Or, you can click here and toggle this.
I typically just use shift tab cuz it's a lot easier and it's a lot faster.
And then, whenever I'm in plan mode, I can start brainstorming.
And not only that, it's useful because Codex will not make changes whenever it's in plan mode, and it also gives Codex to come back to you with follow-up clarifying questions, and I find that super super helpful because we always want to make sure that
Codex knows exactly what it's doing.
We Codex knows exactly what it's doing.
We never the last thing that we want is for Codex to assume something.
We never want it to assume anything.
We always want to make sure that we are on the same page.
And so, let's say I have a command over here going, and I want to build an architecture.
I want to build some kind of a an intricate architecture.
And, you know, I I want to make sure that we built out an architecture.
I may not know exactly how far I want to go, exactly all the details, but I want to make sure that Codex is on the same page.
So, I can say something like, "Come up with an architecture that will allow me to create both web and mobile and desktop versions of the app.
Ensure it is fast and robust.
Okay?
I'm here in plan mode, and typically with this kind of prompt, it's not clear what I want.
I don't even know what I want.
I want Codex to figure it out, but I want Codex to come back to me and confirm with me.
And so, if we execute this prompt, hopefully it's going to ask me some follow-up questions.
Let's see.
And so, as expected, Codex came back with a question, "Which cross-platform strategy should the architecture optimize for?" And there are two questions, okay?
Do we want shared core native shells, and you can click here, and you can get a little detail.
You share TypeScript domain logic with separate Expo mobile, etc.
Do we want a one universal app, or do we want web first wrappers, right?
Because I asked for a cross-platform product, but there's many ways of doing it, and there's many trade-offs with each approach.
And so, it came back to me, and it wants to clarify one universal app, this or that.
So, I'm going to say this one.
I'm going to I'm going to go with recommended.
And now, it wants to know what should the app support for data and reliability.
Offline first cloud sync, cloud only real-time, or local only.
And you can click and you can hover over the little I to get more details.
Users can can work offline and sync across devices.
Cloud only, simpler shared back and live updates, but weaker offline behavior.
Now, here you can also tell Codex something else.
You can say, "Well, I want this, this, this, this."
Right?
Or, you can press escape, and you can ask Codex to explain the question.
I can say something like this, "Explain the last question in simple terms."
Okay?
I don't need to answer it right away.
I can get Codex to kind of walk me through it, right?
Explain the last question.
Let's see what it says.
I was asking two choices.
How should data work?
My recommendation, shared code and native shells plus offline first cloud sync.
If you want this to become a serious app.
So, it tells you these things.
And so, you don't need to pick a pick a choice or tell it something else.
You can close that, and you can ask Codex to clarify.
And now, what we can do is ask me those questions again.
You can do that, or you can simply answer it here.
It doesn't matter.
So, if you do that, it's going to come back with that question here again.
And here are the questions.
I'm going to say this, and now I want to say this because I I I understood what it was asking me.
Got it.
Your choices are that means the architecture should prioritize a shared TypeScript app a brand used by this, this, this, this.
And that's good.
And now, I can exit the plan mode, and I can say, "Go ahead and do it.
Go and implement this." Something like this or something more specific.
And so, plan mode is something that I use all the time, and you should definitely make it an integral part of your workflow as well.
Now, I want to talk to you about some advanced Codex workflow techniques.
These are the little tips and tricks that I use all the time and will greatly help you when using Codex to to build apps and solve problems.
One thing I do all the time is asking Codex to create a handoff before I start a new thread.
This is useful when a thread has gotten long, when the context has been compacted, or when I want to move from planning to implementation with a cleaner workspace.
A good handoff is not a vague summary.
I want the current state, the files that changed, the commands that were run, what still needs to happen, and any risks or open decisions.
That way, the next Codex thread can pick up the work without rereading the entire conversation or repeating mistakes.
And there are two main reasons that you may want to do that.
The first reason is that after a while, this conversation here, this it starts to degrade.
As you keep typing new prompts, and as the model answers, the convo starts to degrade.
Now, Codex tells you that it automatically compacts its context, and that is true.
You do not need to start a new convo every time.
It it does automatically compact it.
But, you're going to notice that as your conversations get long and you keep asking questions and it keeps answering, quality of the responses will degrade.
And moreover, you know, your conversation is going to get fairly long, and that might, you know, degrade the performance of the app itself. right there is a good use case to create a brand new convo using the handoff as the
starting point.
Another really, really good reason to create a new convo with a handoff is when you're trying to solve a problem but the model is not getting you the solution that you want.
And so if you're trying to solve a particular problem and you've prompted the model a couple of ways and it's trying it's giving you answers but those answers are not ideal.
It doesn't give you the right algorithm.
It doesn't give you the right solution.
There's a high chance that each additional prompt is not going to get you closer to a solution.
So when you experience something like that the best thing that you can do would be starting over by creating a new conversation and giving the model a handoff with the things that it has done and the things that that are not working
and where you are right now and telling the model as part of that handoff prompt to work on a different solution.
And because AI is is can be fairly random in some ways, creating a new conversation usually give you a different solution.
That solution might actually work.
Now there are various ways to create this handoff prompt that's going to give you a handoff for a new conversation.
One of the simplest ways would be by typing something like this.
Okay, this is a very very simple prompt that still delivers a fairly high quality handoff.
Okay, so you can type create a detailed handoff prompt that I can use for a new Codex convo.
Okay.
So let's go ahead and send this and see what what answer we get.
Right.
And this is the response we got.
I want to design a fast robust cross platform project management app.
So it's essentially summarizing everything that has happened in the convo, right?
This is the context.
This is the product direction.
This is the architecture preference.
Please produce a detailed architecture plan covering all this important, right?
Also some constraints.
And so what you can do is you can copy that and then you can go ahead and create a brand new convo.
For instance, you can click here and you can simply paste it.
Okay, well you you don't want to delete that.
You want to delete like the top.
You want to delete that.
And when you paste it, that's going to create a conversation and the model can simply continue where the other conversation.
That is one way to do it.
Another way to do it is by giving it a more detailed handoff prompt.
So, let's go back here and let's go ahead and give it a slightly different prompt.
Create a clean handoff for a new Codex thread.
Include all of this.
Do not make any changes.
So, this prompt is slightly more detailed and you can bet that it's going to give you a higher quality handoff.
Okay, so let's go ahead and run this and then we can compare the two handoffs and see which one is more detailed.
Now, take a look at this handoff right here.
Look at how much more detailed this handoff versus this handoff right here.
Now, this handoff would definitely work, but this handoff would work even better.
Going to result in a higher quality conversation.
Okay, so given an option, you can certainly use a prompt such as this, but it's a lot better to use something like this that's more detailed and it's going to save you a lot of time in the long run.
The next practical habit is using Git as a checkpoint system.
Before I let Codex do meaningful work, I want to know whether the current state is clean.
After each milestone, I want a diff that is small enough to review and explain.
That is what keeps an agentic workflow from turning into one giant mystery edit.
The useful trick is to make Codex summarize the diff in plain English.
I do not only want to know what files change, I want to know why they changed, whether each change belongs to the task, and whether anything looks unrelated before I accept it or move on.
So, when you are working in Codex, you do not want to be just creating edits without recording what the edit is all about.
And in software development, you want to be using a source control system, such as Git.
We have here a convo where we were building an app, and in this particular instance, we did not use Git.
And so, I do not know, I have no way to check which changes were made where.
I have no way to revert the changes.
I have no way to see what the actual change represents.
And if working in, you know, bigger teams, I do not know who was the person responsible for the change.
Now, in this convo, we have another problem, okay?
The current working directory is missing.
And so, in this case, what we can do is we can take a handoff prompt, and we can create a brand new project, insert that handoff prompt, start a new convo, and then enabled Git, and set up so that we have Git control.
So, I'm going to go over here.
I'm just going to do start from scratch.
I'm going to paste this handoff prompt to start a brand new convo.
And I'm going to preface it by telling it to create a Git repository.
Create a Git repo if it doesn't exist, and do an initial commit with the code.
So, I'm going to run that.
And the idea is we are going to initialize Git if needed, and then commit the current code as is.
All right.
So, what it did was it initialized a Git repo in this directory, and made an initial commit.
Initial architecture plan here, okay?
The workspace was empty when I inspected.
There was nothing there.
We just created a brand new project, and it essentially created an architecture.
And so, at this point, we have here a Git repo, which means that each additional change will be automatically tracked.
So, for instance, I can say something like, "Make the smallest possible change, a route that displays 'Hello, world.'" Okay?
Let's go ahead and run that.
So, it now created this index.html.
We're going to open it, and it says hello world.
And so, now what we can do is we can say check the Git status.
And obviously, this file is not going to be in our Git, right?
You see this is not tracked, an untracked file.
So, what we want is we want to commit index.html as initial commit.
Because in Git, you can have files that are tracked, and also files that are not tracked.
The files that are not tracked are the files you don't really care about, right?
You don't You don't really need to store them in Git.
These could be like media files, or generated files, code generated files that that can always be regenerated.
But for your actual code, you always want to be tracking them, right?
So, we we have this commit, and now this files is going to be tracked, and it's going to be committed as well.
Okay?
So, now it says committed as initial commit, right?
We have this It's now on main, which is the a type of branch.
And so, now if we go into changes, you can see that we can select a branch here.
And as you can see here, it says no file changes yet.
And that is because we have not made any changes since tracking that file.
So, what I'm going to do now is that make index.html output hello James.
That has updated, and now it says hello James.
And now what we can do is we can commit it.
Commit index.html.
And now if we type give me Git status, it's going to tell you that it's not It hasn't been committed, right?
It's modified, and it's tracked, but it hasn't been committed.
Okay?
It has been modified and uncommitted, right?
You see this M, which means this has been modified.
So, we're going to do commit index.html.
And it will also generate a default message.
Here it has committed it.
We can click here, and we can click select review, and you can select the the of change you want to see.
You can see on stage, stage, branch, or last turn.
And when you click on last turn, you can see what was changed, okay?
Because we are we are using source control.
We can see the stuff that was changed.
As you can see, this is the one line that was changed.
Red means it was deleted, and green means it was inserted.
So, this was deleted, and this was inserted.
We basically changed the message that it says from Hello World to Hello James.
And so, Git is very very important.
If you're working on any kind of project, right?
Anything beyond something super simple, you want to create a repo, and you want to implement source control into your workflow.
One advanced trick that becomes more useful as projects get bigger is working with Git work trees.
A work tree is a separate working copy of the same repo connected to a different branch.
The practical benefit is that you can keep one version of the app stable while trying a risky feature, new eye direction, or refactor somewhere else.
This matters a lot with the genetic coding.
If you want to try two approaches in parallel, or or keep a clean version ready for recording while another code expert keeps building, a work tree gives you that separation.
Now, what's nice about work trees is that they give you parallelism.
So, if you are working on multiple features in your app, maybe you're using sub agents or something else, it's important that no two processes work on the same file at the same time, because that's going to create corruption.
And so, what you typically end up doing is you create a work tree for a specific feature, and you work on that work tree.
And when you're done, you merge the changes back onto your main tree.
Now, that doesn't mean you can only create one work tree.
You can create multiple work trees, as many as you want, and you can also work on the same feature because after one worktree is done, it's going to merge to main, and if there's a conflict, it's going to resolve that conflict, and then
another worktree, when it is done, it's going to also merge into your main branch where your development is happening, and if there's a conflict because they ended up working on the same feature, it will also resolve that
conflict so that there's no corruption.
Now, this is a feature I use all the time in big projects.
Now, let's say I'm working on this app over here.
You can ask Codox to work in main without creating any worktrees, or I can ask it to create a new worktree while it makes changes.
So, let's continue what we were doing.
So, let's say I want to change this index.html, and I wanted to say something else.
So, I'm going to go ahead and tell it, "Make index.html output hello, plan." But this time, we want to make sure that it works in a different worktree.
So, what I'm going to do is I'm going to say, "Work in a new worktree, and then merge, clean up after you're done, okay?" So, we're going to go ahead and run this, and now it's not going to modify the file on our main branch.
It's going to branch out, do the work, and then come back and merge the changes back in.
And so, as you can see here, there's more stuff happening.
I'll make this in a separate Git worktree, commit there, fast forward merge changes to main, and then remove the worktree and delete the branch.
And that's exactly what it did, except it did it very, very quickly because this was a, you know, a very small change, but you saw it doing all that, and before you know it, it did all that, made the change, merged everything back,
and deleted the worktree.
And so, that was a very, very small change.
Now, another thing that you can do is you can click here, and you can fork this whole convo in a separate worktree, right?
So, if I click over here, that will fork it into a separate work tree.
And here you can see we're working in a separate work tree here, okay?
So, now as we're doing work, we don't need to constantly tell Cortex, "Hey, you know, create a new work tree, clean up, and do all this when you're done." We can continue working and then merge our changes.
So, for instance, I can say something like to say, "Hello, USA." I can simply do that.
I don't need to tell it to work in a separate work tree because we are isolated here, okay?
So, here's the file, we can open it.
Hello, USA here.
And so, at this point, we have merged the work tree change into main and cleaned up the temporary work tree, right?
Index now displays Hello, USA, and we no longer have this temporary work tree.
We've done everything that we needed to do separate from the main branch and then we merge.
And as far as we are concerned, this convo is now finished because it's tied to the old work tree, but we've did everything we needed to do.
We've updated the files, we merged back into main, and so now we can start a brand new convo to work on a different feature.
And so, this is really nice for feature isolation.
Create a new work tree, work there, merge back into main, and then simply click here to archive this conversation.
Another advanced trick is using sub-agents for parallel work.
The idea is not to throw a whole project at a bunch of agents and hope it works.
The useful version is much more controlled.
An agent explores the code base, another works on a clearly bounded part of the UI, another checks tests, or reviews the result.
The key is ownership.
If you use sub-agents, you want each one to have a specific job, clear file area, and a clear output.
That keeps the work parallel without turning it into a merge mess.
Sub-agents are best when the tasks are independent enough to run side by side, but still simple enough to integrate at the end.
Now, sub-agents is another really cool feature that will help speed up your workflows.
And this is something that I use all the time, especially in big repos, in big codebases, to do multiple tasks in parallel.
All right, so here I have a very, very simple Kanban app.
If I click over here, I have this.
I can drag and drop. create new tasks, etc., etc.
And so, what I want to know is if I'm going to be building more and more features, complex stuff, new functionality, I want to use sub-agents to help me kind of make sense of it, you know, and help me build it in the most efficient way
possible.
And so, I'm going to paste this prompt here.
Plan how to split this feature across sub-agents, right?
Do not edit files yet.
Recommend.
We're going to jump into plan mode.
Preface this with the feature.
Add to party apps, such as make.com, Zapier, etc.
And so, I'm going to I'm going to leave it at this.
And I'm going to go ahead and run this.
And so, let's go ahead and run this.
And here we have the plan.
The safest split is for the main thread to own architecture, shared data contracts.
Sub-agents can work on isolated feature areas inside index.html, but they must not all edit the same broad sections at once.
Okay, so we have main thread responsibilities.
And then we have sub-agent delegation.
Integration settings, integration data model, web hook event dispatcher.
So, I'm going to copy this plan, and then I'm going to preface it by telling it spawn sub-agents right away to help design the third-party API functionality.
Let's go ahead and run this.
Right away, it's telling me, "I'll spawn the four-focused sub-agents and keep the main thread responsible for the final." So, as you can see, it's spawning them.
And each sub-agent gets a prompt.
It's basically like me creating a new conversation.
Your sub agent one integration, your sub agent two integration data model, your sub agent three, and they're working independently.
And here it tells you all four sub agents are running.
While they work, I'll inspect the current partial edit, and you can see these background agents here.
Okay?
So, they're working in parallel.
They're working on their own thing.
They own different pieces of the app, and they're kind of working.
And when they're done, they come back to this main agent, right?
This combo here, and tell it what happened, right?
They come back with the result.
And so, now it says the persistent sub agent returned a cleaner data shape.
And here you can actually click on any one of these and see what's happening.
You can click here, you can come back, you can click on this one, and see what it's doing.
So, this one is finished.
Some are finished.
How about this one?
This one is also finished.
And so, that's how it works.
And when they finish, they come back, and then it decides what to do later.
Maybe it's going to respond, and maybe it's going to spawn something else.
Maybe it's done.
They did what they needed to do.
They finished their isolated task, and their services are no longer needed.
And again, sub agent is something that I use all the time.
I use it all the time, especially in big repos, because they help speed up the work by isolating different areas, different responsibility clusters, to different agents.
And that's a huge productivity boost.
Now, in this section, we're going to bring all of the previous sections together by giving Codex project instructions in an agents.md file.
This is where you can write down the rules you want Codex to follow every time it works in a specific repo.
Which commands to run, which folders matter, which patterns to preserve, how to verify changes, and when to use work trees or sub agents.
This is useful because you do not want to repeat the same project rules in every prompt.
If this project should always run a specific check, use a specific design system, avoid certain files, create work trees for risky experiments, or split large tasks into sub-agents, those instructions belong in agents.md.
It turns your preferences into reusable project context.
And here's a prompt that you can use, which is going to draft agents.md for this project.
Do not write the file yet include the project purpose, the main folders, work tree, sub-agents, stuff like that.
So, we're going to jump into plan mode.
We're going to go ahead and run that.
And here's the draft for this agents.md tailored to the current project, right?
Project purpose, main folders, verification steps after changes.
Serve the app locally, coding and design patterns to preserve.
When codex should use a work tree.
Use a separate work tree when these happen.
When codex should use sub-agents.
Use sub-agents when work can be split into clearly separate ownership areas such as these.
What codex should avoid changing without asking?
All of this.
Now, the beautiful thing about this agents.md file is that you no longer have to worry about, you know, asking codex to spawn sub-agents or create separate work trees.
It will be automatic.
It's like saving you from typing those prompts every single time.
So, if you want to add a feature, well, you don't need to tell it, "Hey, create a new work tree."
Or make sure you spawn sub-agents.
It's going to be automatic.
And so, at this point, you want to review this proposal here, and you want to save this as agents.md.
So, I'm going to say, "Save it as agents." So, now it's saved it as agents.md.
You You can click here, and you can see the whole file right here.
And this file will be used only for this specific project.
They're based on a repo.
So now for my next prompt, if it's a simple change, it's not going to create a work tree probably.
And it will also not spawn sub-agents, but if it's a more complex change, it should probably create a work tree and maybe even spawn sub-agents.
So let's go ahead and see if this works.
So now, let's go ahead and ask it make a simple change that that changes the title for the main index.html to awesome Kanban app.
Okay, I'm going to go ahead and run it directly.
So now the title is awesome Kanban app.
Okay, but if we ask it something more complex, like, you know, figure out a third-party, you know, integration, design, you know, support API calls, or support MCP, or anything like that, it should probably use sub-agents to figure it out and
implement the changes in its own work tree.
So let's see if this works.
Add a robust API functionality for handling outgoing API calls and incoming webhooks.
Okay, I'm going to jump into plan mode to see if it has any questions for me.
See how it says agents.mdx explicitly says not to add a backend or proxy without asking.
Okay, so it did consult this.
So we're going to add a backend.
We're going to create tasks and let's see if it's going to create a work tree automatically and maybe even use sub-agents.
Another question, should the new backend become the source of truth for Kanban tasks?
Yes, let's just go with the recommended.
Okay, we're going to expand this plan and we're going to see if it's going to be using a work tree or sub-agents for any of these things.
Okay, so I do not see it's going to be using sub-agents.
Update.
So before we run it, let's ask it, do you plan on using a work tree or sub-agents for this task.
There you go.
Yes, for implementation I'd use a work tree because this changes the project from a static only app into a node server app with API routes and persistent states.
I'd also use subagents, but only for bounded design review tasks.
And so it's telling you right there that it will be using a work tree and it will be using subagents, and you can bet that this would never happen if we didn't have this agents.md file.
And so agents.md is super important.
It's going to save you a lot of time because it basically codifies repeatable tasks.
So you never have to re-prompt the same thing over and over again.
Now, before we build real apps, there is one security habit you want to establish early.
Codex can help inspect configuration, but we should be careful with secrets.
An environment variable might be safe and public, like a front-end URL, or it might be private, like an API key or deploy token.
These should not be casually pasted into a prompt or committed into the repo.
The other important thing is that local config and deployment config are not the same thing.
A local.env file helps the app run on my machine, but Vercel does not automatically copy that file into production.
For a deployed app, the required variables need to exist in the deployment environment, too.
So the right workflow is to ask Codex to identify what variables are needed, classify them as public or private, and then advise you where each one needs to be configured.
Now, let's say we're building on this app, and we are adding external services.
So one of those services might be something like OpenAI, or another service could be Open Router.
Well, every time we connect to external services, we need to connect them using an API key, and those API keys are needed for those services to track how much usage we have so that they can bill us accordingly.
And so in your app, you need to be storing those API keys because when you're going to be making those requests, regardless whether that's an API token or some kind of deployment token or pretty much anything else, you need to be very, very careful
how you are storing them.
And so in this app, let's say we want to use OpenAI here.
Here.
So I'm going to say API stack and use the following key.
And let's say we give it a sample API key.
Sample API key as a sample.
We're going to jump into plan mode.
We're going to select this API stack only.
We're going to be doing native fetch.
All right.
So it went ahead and created our first API stack here.
And right here it mentions this file where it stores this key.
And so if you want to use a real key, you never want to paste it in chat.
What you want to do is you want to open this file and modify it directly.
So that's what we're going to do.
We're going to open this in the editor here and I'm going to paste this key right here.
And this right here is my actual key.
And so now I have an OpenAI, have an API stack and I have a real key in this file here, okay?
And now what I need to do is I need to run a check and ensure that we are storing the keys the right way and, you know, they're going to be deployed correctly and all that.
And so I have this prompt to do just that and it says here inspect this project for environment variables and deployment config.
Do not print secret values and then tell me all of these things.
So we're going to run this.
So now it's telling us that environment inspection required environment variables required only for post OpenAI front end variables, right?
We don't have any public because, you know, sometimes your app can have both private keys and public keys, right?
So if you're using something like Superbase, Superbase has an anon key or a secret key or they call it something else now.
And the idea is the public key can be made available, but it has very, very strict permissions.
Whereas something like your OpenAI API key or your open router API key or something of that sort needs to be very, very secret, right?
Because with that key, you can make a lot of calls and do a lot of things.
And so here it says we have a private key.
OpenAI PK is private and must stay server-side.
It is used here, needed locally.
Hosted service config or cell or another node capable host config as a private server environment variables, port, etc., etc.
Okay, dot env examples cuz currently untracked.
So, what we have safest next step before deploying, replace it with only this.
Only config the real key in the deployment provider's private environment variable settings.
Okay.
So, I want to store the key locally for testing, and I want to ensure that it will be deployed also later.
And so we want that key locally so that we can test, but we also want it to deploy it.
And that means we need to create a couple of files, one for local settings and one for deployment settings.
And it's telling us which is the wrong file for local secret because examples are usually committed.
So let's say we're doing a Vercel deploy.
We're going to switch that.
And there we have it, store the real key locally in an untracked.
Right?
We need to make sure that this is never committed to Git.
It needs to be untracked.
So we're going to expand this, submit this.
Do not use public.
These are not public keys.
Replace dot env example with create dot env.
Okay, so it's it's going to be creating this dot env.
The secret is now ignored in dot env.
And then I'm updating readme and agents to make the local and Vercel setup unambiguous.
And that way we have the local key stored for local testing.
It's never going to be committed into Git and then we'll be able to send that key, also deploy that key on Vercel because Vercel is going to be running our app.
It needs access to that key as well.
It needs to be secret.
Okay, implemented the local Vercel key setup.
And so in this point we've made the the secret key management and not just secret key, any kind of key management is part of our workflow and we can be assured that we have the key for local stuff and we also going to be
deploying the key correctly for production stuff, regardless what kind of host you're using.
Now, when you're developing an app, you're going to be writing code for the core stuff, but a lot of things that your app will be doing will be done by existing code.
You're not going to be building your app from scratch, right?
You're going to be using existing libraries, existing frameworks, stuff like that.
And those are called dependencies and that is because your app depends on them to function.
Without those things, your app will not function.
And you need to ensure that you have a process for managing these dependencies.
And I have here a prompt that you can use that will help you inspect this project before adding any dependencies.
Do not install anything yet.
Tell me all of this.
And so let's go ahead and run this prompt.
Right, and this is what it came back with.
Dependency project inspection.
Framework and package manager.
Framework non-detected.
This appears to be a plain static JavaScript app served by a custom node server.
Now, of course, this will definitely not be the case when it comes to your real apps.
You're going to be using some kind of framework, something that will make your app behave like a real app.
Here we just have a very, very plain app.
Package manager non-detected.
Oh, none of this.
And of course, when you're going to be building your app, you're going to be having all this as well.
Available scripts.
No NPM scripts exist because there's no package manager, because there's no manager, because there's no package.json.
Current documented run command, we just have this very, very basic.
Definitely not the case when it comes to production apps.
OpenAI local env run command.
So, we have this relevant install dependency.
So, we do have some basic node built-ins.
All of these OpenAI calls native fetch, okay?
So, it's native.
It's not using the OpenAI SDK, which would have been a dependency.
Can the task be completed without adding a package?
Yes, for the current OpenAI API stack and local serving, the project already calls OpenAI via native fetch.
If a new package is recommended, no package is recommended right now.
A package could be recommended later if we use Express, Fastify, we start using the official SDK, .env files likely to change if proceeding, right?
API behavior, OpenAI calls, front end, etc., etc.
And so, dependencies are important, and you want to limit your dependencies.
You don't want the system to just add dependencies blindly.
And so, this is going to be super useful as you're building your app so that you can streamline it.
Because as I was just building this simple app, Codex asked me whether I want to use OpenAI SDK or native fetch.
And you always want to err on the side of native stuff.
If you can avoid a dependency or an SDK, that would be the right move.
And that's exactly what I did here.
I just used native, right?
I could have used an SDK, but I decided I'm just going to keep it as simple as possible.
And that's really the goal here.
And this prompt and making this part of your workflow is going to enable you to keep your app streamlined so that you have less dependencies, less chance for something to go wrong, less bloat, etc., etc.
So, definitely add this to your app building workflow.
And then ensure that Codex knows about it by making this a part of your agent's MD or a separate step.
Now that the Cortex foundation is clear, we are going to start with something intentionally small.
The first step is called Habit Card.
And it is a tiny habit tracker where users can add habits, check them off for the day, and see a simple streak or count.
This is a great starter project because the result is visual, easy to check, and small enough that every step stays clear.
Cortex can plan the app, build the UI, show us the file changes, open the browser, polish the result, and get the front end ready to deploy without also introducing back end complexity.
Habit Card gives us the core app building pieces without much product complexity.
A form, a list, a daily check-in state, simple feedback, and later local storage persistence.
It does not need authentication, payments, APIs, or a real database at first.
This first app is about getting comfortable with the full Cortex loop on something we can understand immediately.
Now, here in Cortex, the first thing that you want to do is you want to click here, and you can select start from scratch, but I like to have it set to a specific folder.
So, I'm going to click on use an existing folder, and I have here a folder prepared called demos.
I'm going to open this here, and now we have a new project that has been associated with a specific folder.
And here we can start a brand new And here we can start a brand new conversation for the app that we are building.
Now, anytime I start building a brand new app, I always want to create a visual architecture.
Being a visual person, this gives me an understanding of the scope of an app, how the app looks, the user flow, and stuff like that.
And we can actually use Codex to create that visual architecture for us.
Now, I have here a prepared prompt that creates this visual architecture, and we want to ensure that we use the built-in image generation that we have as well as Chat GPT to do it.
So, I'm going to press forward slash, and I'm going to search for Imagen.
Okay, and then we can say something like use Imagen for the following.
And I'm going to go jump into plan mode, and I'm going to go ahead and and execute it.
Okay, and it came back with a habit card architecture diagram, and here it says create one raster infographic using this.
So, everything looks good.
There there's some diagram content, visual direction.
That's a wide composition.
Looks good.
We're going to say yes, implement this.
Okay, and that is our habit card.
That is our visual interface.
And right away, I can see what this app is all about, right?
We have add habit, we have habit list, we have today check-in, and we have a streak count.
Then we have an empty state, we have local storage later, and we have the Versel deploy.
And so, this is a very, very simplified visual architecture, but it is very, very important.
I want to see right away what I'm dealing with, what I'm working with, how many pages, you know, how complex is the app, how simple is the app.
I see that this is not going to be, you know, a very complex app.
It's going to be a fairly simple app, but we do have some interesting bits here, right?
This empty state, we have some local storage, we have a Versel deploy.
And so, it's a fairly simple app, but still a fairly workable app.
Now, the next thing that I like to do is create the entity architecture.
And this is where we're going to be mapping the small set of entities behind the app.
So, even though the first version is front end only, it is useful to name the habit, the daily check-in, and the streak count summary before we add a local storage later.
And this right here is the entity diagram.
And I'm just going to say use ImageGen for below.
Going to go ahead and run that.
And I can run it directly.
I don't really need to run it in plan mode because it's just generating a diagram.
I'm not asking it a complex, ambiguous coding question or giving it a an ambiguous coding instruction.
And here is our entity architecture.
And unlike the visual architecture that we've generated previously, here we're actually seeing the entities that will later become the actual storage schema.
Okay?
So, we have the habit, we have a habit check-in, we have a streak summary, and then local storage as the back end.
So, no back end, everything is stored locally with local storage.
And this is also very, very useful because unlike the visual diagram here, I want to see the actual schema.
I want to see what kind of, you know, elements or what kind of, you know, tables we might need to translate this app schema into.
And as you can see, this is fairly simple.
Habit, habit check-in, streak summary.
We have these three main functions here, these three main features that were translated into entities.
Okay?
And then we have And for each entity, we have things like name, created at date, completed count.
Now, at this point, we have both the visual architecture as well as the entity architecture.
And this is something that I like to do for pretty much every single app that I build because I want to get an idea what I'm really dealing with, right?
I want to have an understanding of the the scope of the app.
And right now I can see exactly where we are.
I can see that this is a fairly simple app.
I can see the entities.
I I I see the app building.
I see the future of the app.
I see the app coming together.
Now, the next thing that we want to do is we want to use the plan mode to ask Codex what stack makes sense for this first app.
The important part is that Codex should recommend the simplest path before it edits any files.
And as you might have guessed, we are going to be doing this in plan mode.
All right?
And here's the prompt that we're going to be using.
Do not edit files yet.
Inspect the current project and recommend the simplest stack for Habit Card, right?
And we ask it what what it should come back with, and we're going to press shift tab to enable the plan mode, of course, because we want it to come back to us if it needs to and ask any clarifying questions.
We need it to clarify everything before we go deeper into the building the app.
Okay, I'm going to go ahead and execute this.
And let's see what it comes back with.
And in this case, it did not come back with any clarifying questions.
Okay?
So, it's writing a plan.
And the recommended front end stack is going to be React TypeScript, right?
It wants to use Tailwind CSS for fast simple styling.
This is the testing library.
No React Router.
It's a simple single screen app.
No global state library.
Okay?
Back end decision.
No back end is needed for the first version, right?
And that's what we told it.
We said bias toward a front end only implementation first.
And then it's also going to use browser verification.
Okay, open the app in the browser, verify all this.
Okay, so let's go ahead and and run this.
All right, so it worked for about 10 minutes and it has implemented the habit card as a front-end only React TypeScript app.
This is where it is and this is what it did, right?
It built the app shell and did, right?
It built the app shell and UI, habit logic, local storage, components for adding form for the add form, list card, and empty state.
Tests for app flow logic and storage, and verification passed nine tests, production build succeeded.
And it also gave us a dev server that is running the app.
And so this dev server is running on my machine right now and this allows us to quickly test things without needing to deploy and test there.
And on the right-hand side we have this card that is always available and it shows us things like artifacts, background terminal, and basically things that were created as part of this conversation.
Okay, so we have the images that were generated, we have the index.html for the app.
We have the background terminal that is running here.
And so I can click here and that's going to go ahead and and go to the server that's running our app.
So let's do that.
That's going to go ahead and open.
It's also going to open the browser right here in the right-hand pane.
Now, before we had we had no right-hand pane, but if you click here, you can see that it opens up here right there.
Another thing that you can do is you can open the browser directly.
This is the front page here and I can add a habit, habit one.
I can click here and we have a habit here.
I can add another we have a habit here.
I can add another habit, habit two or maybe gym.
And it's creating these these habits here.
So I can check it off and that is done.
Check it off.
And so it's a very very simple app.
And if you want to see the actual files that were created as part of this, you can simply right click here and open in finder.
And this is the React app that was created for this here.
Once the UI works, we'll add a small persistence step with local storage.
Habits and today's check-ins can persist after refresh.
So, the app starts to feel real without introducing a backend yet.
And this is the prompt that we're going to be using.
Upgrade habit card so habits and today's check-ins persist with local storage.
I'm going to jump into the plan mode here just in case.
And I'm going to go ahead and run this.
Now, it's actually telling us that inspection shows that habit card already implements the requested front-end only local storage persistence.
Okay?
So, no file edits are needed for this upgrade.
So, we must have done this in one of the previous prompts.
But, if we hadn't done this, then we would run this prompt.
But, still, let's go ahead and run it.
And now it tells us that persistence works through this storage using these keys and a lot of other details here.
Okay?
So, we can reload this and now we have something called local storage.
And what this means is that this app does not have a backend, but this stuff is going to be stored on my browser.
Okay?
So, if I'm using this instance, it's going to be stored on this browser.
If I run it in a different browser, then it's not going to have the data.
So, whatever I do in that new browser is going to have fresh data.
And so, this is a way to create, you know, storage just for testing me.
I mean, you can use this in production as well for certain apps, for certain situations.
It does have its place, but it's very, very limiting compared to something like Supabase or something of that sort.
And so, we can delete all this, we can do all that, we can start.
And if I copy this, run this app, and I load this app, there's nothing there because I'm using a brand new browser instance, and I can create, I can do, I can check off, I can do this, I can do all that, but if I go back to CodeX, there's nothing here,
right?
It's not syncing, it's not using my account, and it's not using a unified back-end, so that no matter which browser I use, it will still store the data and keep track of everything.
That is not happening.
Once the app works locally, it's time to deploy it.
Even though Habit Card is simple, this is the first moment where CodeX helps move something from local code to a real-life URL.
And so, CodeX can do the whole deployment process.
It can run the local checks, prepare the projects, connect or create the Vercel project, deploy the front-end, and then verify that the live page actually loads.
And additionally, you want that first deployment to feel very concrete.
We're not just writing code and hoping it works.
We're taking a small app, shipping it, opening the deployed URL, and checking the result like we would with a real project.
And here is my Vercel deployment prompt.
Prepare Habit Card for deployment on Vercel, and verify it works as a small front-end only app.
So, at this point, we are just preparing it for deployment.
We're not yet deploying it, but still, this is an important step because this will ensure that once we do deploy it, that it will work without any issues.
I'm going to go jump into the plan mode, and remember, we do have that Vercel plugin enabled, and so it should use it automatically with no issues.
I'm going to go ahead and execute this.
All right, so it went ahead and created a Vercel deployment plan.
Okay, Habit Card is already a suitable for a front-end only for sale deployment, okay?
And this is what it's going to do, and here is the deployment implementation.
Keep the app front-end only, do not add Convex, which is a back-end APIs or database server functions, or external services, none of functions, or external services, none of that.
And local storage persistence is expected to be local to each browser and deployment origin.
Data will not sync across devices or URLs, okay?
Let's go ahead and implement this plan.
And 3 minutes later, it tells us that deployed Habit Card to Vercel and verified it live, okay?
So, this is the live URL.
We can click here, and that's going to open our Habit Card app running on the web, right?
Accessible to everybody.
And this again uses local storage, okay?
So, we can do that over here, and of course, you know, it's not going to be available in another instance because it's all stored locally.
And so, at this point, we designed, architected, built, tested, and then deployed a very, very simple web app that uses local storage as our back-end, and this proved the whole cycle for us.
Now, in the next several modules, we're going to be doing the same, but with more complex apps, slowly adding things like authentication, back-end, and other important components that you need for a real production app.
The second app we'll be building will be a mini team voting app where people submit lunch or snack ideas and vote on the winner.
It's a bit more complex with more features, but still fairly simple overall.
Snack Vote has a submit option form, a snack option list, vote buttons, a live leaderboard, and a winner card.
This app will serve as a bridge between a simple front end and the larger marketplace app we'll build right after this.
Instead of jumping straight into complex entities such as customers, bookings, and permissions, we can first learn the back end pattern on a much smaller app.
Snack Vote only needs snack options and votes.
Someone can submit an option, other people can vote, and the leaderboard updates as the vote counts change.
That gives us enough structure to understand live back end data without making the product itself hard to follow.
As always, in this first step, we'll be mapping the Snack Vote interface visually.
The app should be simple enough to understand immediately.
Create or join a voting room, submit a snack option, vote once, watch the leaderboard update, and see a leaderboard update, and see a deadline-driven winner state.
Now, here in Codex, we're going to come here, and we're going to create a new project from an existing folder.
We're going to click over here.
Here, we're going to create a new folder.
I'm going to type Snack Vote, and we're going to click open.
And now, we have a brand new project configured here in Codex.
And here's our prompt for the visual architecture.
We want to preface that we want to use this image, and we're going to jump into plan mode and execute this.
And that looks good.
We're going to go ahead and expand that.
Go ahead and implement that.
And here's our Snack Vote visual architecture.
Look at this beautiful diagram.
So, you can create or join a room.
In the current room, you're going to be seeing this, members and the time left.
You can You have here a snack list.
You have your your voting buttons.
You have a live leaderboard, and you have a closed winner state.
Okay, with some voting deadlines here, here as well, and then we also have our Vercel deploy.
And so this looks like a fairly fun app with a lot of things going on, and so I can't wait to implement this.
And we're also going to be adding more and more features as we go along.
Next, I want to go ahead and create an entity diagram, an entity architecture diagram.
So here's our prompt.
We're going to do this, and we're going to go ahead and execute that.
And here's our entity diagram.
Okay, so a room contains multiple snack options and has one voting session, right?
We have the voting room, many snack options, and many votes, right?
We have We have here some winner logic, the winner logic snack option with the highest vote count when now is greater than or equal to closes at, right?
So after the deadline, we have here a legend, and here we have a sample kind of workflow, a pipeline.
Build the UI first, connect it to Convex.
Later, Convex is going to be our backend, and we are going to be leveraging things like real-time updates, persistence, and live vote counts.
And so now that we've generated a visual architecture diagram, as well as our entity diagram, the next thing that we want to do is we want to ask Codex for some stack recommendations, because there are lots of ways that we
can build out this app, and we want Codex to tell us the best options for this specific app.
Here's the prompt that we're going to be using, and I'm going to switch into plan mode here.
And it's telling us do not edit files yet, inspect the current project, and recommend the simplest stack for Snack Poll.
And here's the plan that it came back with.
Here we have the stack recommendation.
We're going to be using React, TypeScript, Tailwind.
And here it says backend later Convex with TypeScript server functions and generated client types.
And so, for the most part, uh this is very, very similar, if not the exact same architecture stack that we use for the first app, except we're going to be adding more and more features, and also connecting a back end for the first time
in this tutorial.
Now, here, instead of clicking submit, I'm going to give it another prompt, and this prompt is going to start building the app, writing the code.
And here's the prompt, "Implement the first working Snack Vote UI without a database or persistence, okay?
Inspect the current project first, then build the smallest useful Snack Vote UI." And there's some verification steps as well.
And we also explicitly tell it to not use convex, local storage, or an API, nothing of that, just keep it to these specs right here.
And we also have some verification.
So, I'm going to go ahead and run this in plan mode, in case it wants to ask me something.
So, it doesn't have any questions for us, and that is because in this specific conversation, we already fleshed out, you know, things like the architecture, the stack, and all of that stuff.
And so, it kind of knows what it needs to do to build out this app.
So, we're going to say, "Yes, implement this." because this is the step where we actually want the code written.
All right, so this has completed, and it says here, "Implemented the first working UI-only Snack Vote app here.
Created all of this.
What works?
Verification completed.
Still needs attention later."
Okay, all of these things, and we'll talk about that a little bit later.
And so, now we have a web preview.
So, we can click on open, and that's going to go ahead and open it right here in our built-in browser here.
This is our app right here.
And so, we have snack and lunch options.
We can create, we can submit an option, add sample snacks, click here, and we have these snack options here with some votes here.
Leaderboard, right?
So, this is winning.
Winner state, current winner leader, spicy ramen bowls.
And so, if I vote on falafel wraps, uh wraps, six.
Right?
So, now it's six.
I can keep voting.
And that's a problem right away, and it's in memory.
And one person can vote multiple times, which is an issue.
We're going to fix that a little bit later, but I can also add a snack.
Let's say sushi.
I don't know, sushi.
Add this, and we have sushi, and I can also vote on that.
So, it's a big big issue, but for the most part, the logic works, the UI works.
Now, once the snack vote UI exists, we're going to replace the mock snack data with real Convex data.
And for those of you that that are not familiar, Convex is a back end that we're going to be using in this tutorial.
It's an awesome back end, super fast.
And if you want to learn more information, I actually did a dedicated video on it that you can find right there.
And so, once we add Convex, our app will stop being a static interface and start behaving like a real database back product.
And the path is very very simple.
Code X creates the Convex schema, adds a query, adds mutations, connects it to the UI, and then we verify the result in both the browser and the Convex dashboard.
And so, when we submit an option in the app, it should be written back to Convex.
When we look inside the Convex dashboard, that snack option should be there.
And when someone votes, the UI should update live so that we can see the front end and back end working together instead of treating the back end like a black box.
And when you head over to convex.dev, you can learn more about this back end.
And this is a great, great back end.
It's very, very fast, and it works really great with agents, right?
So, the back end building blocks for your agents, okay?
So, I'm going to go ahead and log in to my free account.
I'm using the free plan.
I'm not paying for anything.
And this is also an open-source back end, so you can also self-host it if you want.
So, I'm going to go ahead and log in.
All right, so here I am logged in.
And as you can see, I have here two projects.
Now, our app is going to automatically create a brand new project for its data, for its functions, for everything.
And here back in Codex, we're going to paste a new prompt that looks something like this.
Upgrade Snack Votes to Snack Options, and Votes are stored in Convex instead of local in-memory state.
Now, if I press command, comma, and go to MCP Servers, you can see that I do not have the Convex MCP server.
And that is because we're not going to be connecting to Convex via MCP.
We're actually going to be connecting to it using a much simpler method, and that is the CLI, which stands for command line interface.
And actually, before I run this prompt, we're going to check to make sure that we have a connection.
Can you first check that we have the Convex CLI installed, and if not, do install it.
I'm going to go ahead and run it.
Now, I already have Convex CLI installed, but you can just tell Codex to install it.
It's that easy.
And so, it's telling me that NPX, which is a package manager, could fetch Convex temporarily, but the project does not have it installed.
I'm installing Convex into the project now as a dev dependency, okay?
Convex is installed locally now, okay?
I'll verify the CLI results.
And so, you don't need to read lots and lots of documentation, figure out how to connect to it.
You just tell Codex that you want to connect to it.
It's that easy.
So, Convex CLI is installed locally in the project now.
Okay, so now we can paste the prompt.
Upgrade snack votes so snack options and votes are stored in Convex instead of local in-memory state.
So, I'm going to go ahead and do this and let's go into plan mode.
See if it wants to ask me something or clarify something.
And so, it's telling me one material set of choices to open.
There's no existing Convex folder.
So, this app is not linked to a Convex deployment yet.
And that is correct.
And that's what we want to do next.
And now it's asking me which Convex backend should snack vote use for this first backed version.
We want to create a new one, right?
So, we're going to go with the recommended option.
All right, and now we have a new plan to upgrade the existing app from in-memory snack options to live Convex data.
Okay?
So, it's going to do everything and it's going to essentially configure the backend to be Convex, set up the schema, set up the data, everything.
And once it's done, we should be able to go to Convex and see a brand new project here automatically created.
So, I'm going to hit submit here.
All right, so it says it has created a fresh cloud project and synced the backend function successfully.
And now you see this cursor, it's testing out the UI.
And so, right now it's not running, so it should start it out.
And here it says the existing browser tab was sitting on the browser error, so it has restarted it.
And if we jump into Convex, you can see that we have a new project created.
Now, I did authenticate this account with Codex, so it knows that this is my account.
And so, if we click over here, we should see some initial stuff happening, right?
There's some data, functions, and it's not done yet, right?
It's still working.
And while it's working, we can see what has been done.
And it's actually typing it.
Look at that.
Key press input.
It's creating it.
It's testing it right here.
You can actually see it happening.
This snack was submitted through the browser and appeared from Convex.
I'm voting on it.
There is this testing all the flows, which is pretty cool.
All right, now it's telling us implemented the Convex back snack vote upgrade, changed all of this, what changed, verification, edge cases handled, and it is running.
So, we can test it out ourselves.
And so, here's the app.
And now it's Convex here's the app.
And now it's Convex backed.
And so, now if we head to Convex, we should see the data.
So, as you can see, snack options, this has a vote of one.
This is the snack tools, vote one.
And if we increase the vote, it should be two, and it's automatic.
It's absolutely two, and it's automatic.
It's absolutely automatic.
Three, it's three.
It does not need to do anything.
And we can create a new snack option, let's say sushi, submit.
And if we come over here, we see sushi, zero votes.
Now it's winning.
And so, this is essentially the same app, except it's Convex backed.
And we also have functions.
And with Convex, you need to create functions to access your data.
It's not like with other systems where it's automatic.
Here you need list, you need submit, you need a vote, you need functions.
And so, the snack options table or document has three functions.
No files, no schedules, and there's some logs here, so that you can see things are here, so that you can see things are happening.
>> Okay, so this was a fairly successful migration.
We went to no back end to a Convex back end very, very quickly.
Now that snack vote has real Convex data, we can make it feel more like a shared group app.
Instead of one global snackboard, users can create or join a simple voting room like Friday lunch or team snacks.
Each room should own its own snack options, votes, leaderboard, and winner state.
This keeps the app more interesting while teaching a clean back-end concept, scoped data.
We still do not need accounts, permissions, invites, or admin tools.
And here's the prompt that we're going to be using.
Upgrade snack vote so voting happens inside simple rooms.
And I'm going to run this in a plan mode.
All right, so it came back with a plan for us.
This is the room-based snack vote upgrade.
Upgrade the current Convex back app from one global snackboard to simple rooms.
Can I expand that?
Yes, implement that.
Now, as it's running, I am seeing an error message here.
Did you forget to run NPX Convex dev?
Okay, so let's see if it's able to catch that and solve it before the next prompt.
It is done.
So, it has implemented room-based snack vote.
This is what it changed, files, things, verification limits.
Okay, very important.
Limits is actually very, very useful.
No auth, ownership, invites, or duplicate vote guard yet.
And we'll fix some of these a little later.
Room selection, room codes.
Okay, so how does this work?
Let's go ahead and refresh this.
All right, so we can create or join a room, and there are some rooms over here.
So, let's say we type Friday lunch, and so now we've switched the room.
Okay, current room Friday lunch.
And now we can do the same thing.
We're now in in the scope.
It's all about scoping now, right?
And so this is a specific room, and I can say, I don't know, submit.
There we have it in this room.
And now if I switch a room, let's say I switch to launch ideas.
So now it's brand new.
There's nothing nothing there.
And I can say something like I don't know, tacos.
And now we have tacos here.
And if we go into Codex, we should see that.
So we have voting rooms here, and we have different ideas here.
And now if we go into snack options, we can see that they belong to a specific room.
We have the room IDs.
These are the same same room that we have a bunch of different room IDs there.
Probably some of them created by Codex as part of testing, like lunch ideas.
Oh, no, I created lunch ideas.
These ones were created by Codex, these three rooms.
And we can remove them.
I can select these, and we can delete these.
And if we go into functions, we can see that for snack options, it's the same pre-functions.
But for voting room, we have get, create, or join, bunch of other functions that we have.
And And so it pretty much did everything.
It created the functions, it created the schema, it even, you know, created some test data while it was testing.
We'll add one simple rule that makes the voting feel like a real product.
After all, a person should not be able to spam the same room with repeated votes.
And that is because we still do not have real authentication.
And so for this tutorial, a local or session-based voter identity is enough.
And that's what we're going to be doing right now.
And here's my next prompt.
We're going to upgrade snack votes, so each person can only have one active vote per room.
Jump into plan mode.
Run that.
So here's the plan.
One active vote per room.
Add a lightweight session-based voting rule.
Each browser session gets a session ID and Convex stores one active vote per room ID and session ID, right?
So, this needs to be unique.
So, if I voted in that room, my session ID was used to vote.
And so, this needs to be unique.
So, for that room, my session can only happen twice.
If there's a chance for a repeat, it's not going to count it.
Yeah, I can vote in other rooms, but per session per room can only happen once.
We're going to expand this.
Looks good.
And submit this.
So, it is done.
Implemented one active vote per room.
Okay, change this.
What changed?
And it added this get my vote.
Added votes with table.
So, let's go ahead and try this out.
Open this here.
So, let's say I go to another room, friend Friday lunch.
I switch that room and I have my burrito.
So, I'm going to vote for this.
I just voted and I can't vote again.
You see that?
I can't vote again because if I go to Convex and I go into data and I go into votes, I should see a session ID with room ID.
So, let me go ahead and delete all of these so that we can see which one is mine.
And so, now I can vote again.
You see it's back up.
It's it's it's uh reactive.
I'm going to hit vote.
And now you see room ID and session ID.
These two, if they can only appear once in this table.
And so, obviously now it's knows that my session ID and I'm trying to vote for this one and that's not going to work because I'm in this room.
I'm in this room here, Friday lunch.
Friday lunch right there.
Now, if I switch rooms, let's so let's say I go to another room.
Let's say I go into Sunday brunch.
Now, I can vote, right?
I can vote for anything.
I can say tacos.
Submit.
I can vote here.
And let's say I create another one.
I I change my vote.
Yes, I can switch votes, but I cannot pad, you know, with multiple votes an item.
And so, if I go back here, now there are two options, right?
Two different rooms, same session ID, but I cannot vote in the same room more than once.
So, this works really, really nice.
And finally, we'll add a deadline, so each room has a moment when voting closes.
Before the deadline, people can submit and vote.
After the deadline, voting locks and the app shows the winner or a simple tie state.
This adds time-based product state without becoming a scheduling system.
It is still a small group decision app, but now it shows how front-end UI and back-end data can derive open, closed, winner, and tie states from one field.
And here's the prompt that we're going to be using.
I'm going to switch into plan mode, run this.
Oh, it needs a product decision.
Okay, how should users set a room deadline in this small tutorial version?
Quick durations, okay, add simple buttons like 1 minute, 5 minutes, 15 minutes when creating opening a room, date time input, or fixed default.
Let's do quick durations.
That sounds good.
Let's implement this.
And it is complete.
Implemented simple voting deadlines.
This is what was changed.
Added quick deadline choices, no deadline, 1 minute, 5 minutes, 15 minutes.
All of this, and let's go ahead and see how this works.
All right, so here we are.
And let's go ahead and create a room, Friday lunch, and let's say it's 5 minutes.
Select 5 minutes, open the room.
This still says no deadline.
Let's try another room.
Let's say Sunday brunch.
I can select no deadline.
I can do 15 minutes.
Let's Let's 15 minutes.
See, it says no deadline.
Okay, that's a bug.
That's definitely a bug.
So, let's see if it's able to fix it.
That's the first bug that has happened to us as far as I that has happened to us as far as I remember.
And so, we're going to say created tried creating a room and selecting a deadline, but it ended up creating a room with no deadline.
And we don't really need plan mode for this.
There's no, you know, clarifying questions here.
It's a simple bug.
And we don't really want it to clarify.
We we know what the bug is.
We just want it to fix it.
So, there's really no ambiguity there.
It's like It's like I'm not looking to negotiate with Codex.
I know this is a bug and I want it to fix it, right?
There's no other ways.
There's nothing else happening.
Unless I'm wrong, and in that case, it's going to tell me that I'm wrong, that it works, but there's no other paths here as far as I'm concerned.
Now, sometimes Codex would output something like this and it tries to go into plan mode.
And so, right now it thinks it's in plan mode, but it's it actually is not in plan mode.
And so, if I say something like do it, it's going to complain that that I did not uh leave the plan mode while I'm not in the plan mode.
So, this is a bug that I have been experiencing from time to time and it's annoying, but it's it's easily fixable.
You see, it says I can't edit files because this thread is in plan mode, but we're not in plan mode.
And so, the solution is to go into plan mode and type something.
Okay, fix it.
And then it's going to say something like, well, you know, I can't do it.
I'm in plan mode.
And then you can leave plan mode.
And now it understands it.
You see, now it's correct.
Fixed the bug where it's selecting a deadline and opening an existing no deadline room still leaves the room with no deadline.
So, we're going to submit this.
All right.
So, it looks like it fixed it and here it says in the browser, I created a no deadline room, reopened it with 15 minutes, and it showed open plus closes in, etc.
So, let's go ahead and test this out ourselves to ensure that it does really work.
So, we have this and let's do Friday lunch and we said 15 minutes.
And now it shows it.
And let's say 1 minute room one, 1 minute.
There we go.
1 minute.
And let's try no deadline room two, select no deadline, open no deadline.
Okay.
So, this looks like this was finally fixed.
Now, once SnackVote has rooms, session-based voting, and a deadline, we are ready to deploy the front-end app with Vercel and verify that the deployed app points to the right back-end environment.
And this will round up this specific app build.
And here's our prompt, prepare SnackVote for deployment on Vercel and verify it works with the Convex back-end, okay?
I'm just going to do it in plan mode, run that.
Here it's being very, very clever.
One important deployment decision remains, whether the Vercel app should use the existing Convex dev deployment or it should promote it to production, okay?
And this is really, really cool because this is something that I wanted to cover once this was done.
So, if we go into Convex and you click over here, you can see all your deployments, your different deployments.
So, so we we are currently under development deployment, under development cloud.
We can also have a production deployment, preview deployments, other deployments, stuff like that.
So, ideally, once we deploy it with Vercel, we want it to use the production Convex, and that's what we're going to select, because this makes sense.
Production deployment uses the production Convex.
Development deployment uses the development Convex.
And this is our plan prepare snack vote for a production result deploy connected to the Convex production backend.
We're going to hit submit.
And it's going to create a production.
In fact, they already did it, right?
We can switch to this and it's going to have a production environment here created for us.
All right.
So, it has deployed snack vote to Vercel and verified it against Convex production, okay?
So, this is the live URL, what we're seeing here.
Deployment URL and the Convex backend.
Okay, and if we go into Convex, you can see that it's currently pointing to the production URL here.
And we have everything here in production, right?
So, we have the dev instance pointing to the dev deployment in Convex.
And the production instance in Vercel pointing to the production instance here in Convex.
So, we have all the same functions.
We have all the same data.
Everything is there.
The data might be a little different, but the schema is exactly the same and it works, right?
So, we can create a room, whatever new room, 1 minute, open it.
And now 59 seconds, 58 seconds, looks good.
We can create another room, room three, no deadline.
And that has no deadline.
So, I'm currently in this room.
I can create a food option.
I can vote.
I can't vote anymore.
One active vote per room.
This is the leaderboard.
So, it looks really good.
And this time we've actually built a fairly cool and a fun app along with a backend that uses several different deployments depending on where the app is is running.
So, for local deployment, we have the local, the dev convex deployment, and for the production Vercel deployment, we're using the production convex deployment, okay?
So, this was a slightly more complex app.
But, there's still one more app that I want to build, and this app is going to be the most complex of them all.
It's going to bring a lot of these things that we've been talking about so far together, and also we're going to be introducing some new functionality, new features, and a lot of really cool features, and a lot of really cool things.
Now, for this last app in the course, we're going to be building FlexHelp.
And this is where we take the workflow from the first two smaller apps and apply it to a more realistic product, a two-sided marketplace where helpers offer fixers services and customers book them.
The important thing is that we will still build it in controlled steps.
We're not going to ask Codex to build the whole marketplace at once.
We are going to architect the design, and build the UI, then add feature by feature.
And last but not least, we're going to be deploying the app, making it available to anyone.
As always, our first step will be creating the visual architecture.
This gives us the shape of the product, the customer flow, the helper flow, the main screens, and the path from browsing a service to booking help.
Now, here in Codex, we're going to click here, and we're going to create a brand new project.
Now, I like to start it from an existing folder, and I'm going to create a new folder.
FlexHelp.
I'm going to hit create, and I'm going to click open.
And that creates a brand new project.
Now, the first thing that we're going to be doing is we're going to be building a visual architecture, the same thing that we did for the previous two apps.
And this is important because this gives us the shape of the product, the customer flow, the helper flow, the main screens, and the path from browsing a service to booking help.
And here's the prompt that I'm going to be using and I want to specify that we want to use ImageGen.
We're going to jump into plan mode and we're going to run this.
And here we have the Flex Help Visual Architecture.
And here's a better view.
And so here we have the customer, we have the shared marketplace screens, and we have the helper, which are the two sides of this marketplace.
We also have the Convex and the Versel here.
On the customer side, we have browse services, get the service detail, and then send a booking request.
The helper side, we have the helper dashboard and the status updates.
And on the shared marketplace screens, we have explore the marketplace, my bookings, and booking details.
So right off the bat, this app is a bit more complex than the previous apps that we were building.
And that's really the goal, to finish off with a more complex, but also a more realistic, more production-ready app.
Next, we'll map the core entities.
And this is where the product becomes more tangible because we can actually see the real objects in the app before Codex starts writing the schema for code.
And here's our prompt.
We want to preface this with this ImageGen.
All right, and here's our entity architecture.
Okay, so we have the user, that could be a customer or a helper profile.
We have the booking here, booking status, service, category, review, and all of that backed with Convex data.
So at this point, we have a better idea of how the final app is going to look.
I see that we need a user entity and that could be a customer or a helper profile.
And I see that booking is super important because everybody needs to have access to booking, both customers and helpers.
Booking status, service, category, and so right away I see that this is a fairly sophisticated app, but I can see that it's not a super complex or like a super complicated app that you know has a ton of different entities, has lots of things happening.
But still, as we are developing this app, we might be adding various entities and so that will increase the complexity as the development can continues.
Now, the next step is before we start actually building the app is we want to ask Codex to recommend the simplest stack for FlexHelp based on what we need to build based on what we've told it so far.
This keeps the technology choices tied to the product instead of starting with tools first.
And this is super important because we want to start with the blueprint before any code has been written.
And here's the prompt to do that.
Do not edit any files and spec everything.
And I'm going to be running this in plan mode so that I can decide whether I want to accept its recommendations or just leave it in the conversation.
Going to run this here.
This is what it's suggesting.
The workspace is currently empty.
The simplest course-friendly stack is React, TypeScript, Convex for backend data, off, Tailwind for UI, and Vercel for hosting.
Okay, and this is exactly what we've been doing so far.
And and here it's listing some primary routes.
Services, bookings, customer bookings, helper, helper services.
Here we have the data flow, the auth recommendations, UI and component approach, Vercel deployment path, staged build plan, everything.
Okay?
So, I'm going to exit out because this was only a read-only plan mode.
Okay?
I don't want it to do anything just yet because that's what we're going to be doing in the next step.
Now, from the visual architecture, the entity map, and the stack recommendation, we have a pretty good idea of how the FlexHelp app is evolving.
The next thing that we want to do is we want to get Codex to take a first pass at turning that plan into a static product UI.
Now, this first version does not need any back-end data.
The goal is to see how far Codex can get from the architecture alone to viable UI mock-ups.
And this is the prompt that I'm going to be using.
Build the first static FlexHelp UI design from the existing architecture and stack direction.
We're going to stay in the plan mode.
Going to go ahead and execute that.
And let's see how well it does.
And here's the plan that it came up with.
Okay, the goal is a polished visual prototype that matches the FlexHelp architecture.
Okay, and here it lists some details about the architecture.
And let's go ahead and run this prompt.
So, it finished working.
And let's go ahead and see what we have here.
Okay, so this is our app.
And this is kind of how it looks like.
Okay, book practical help.
And we have some active requests.
We have booking entry, helper dashboard.
We have this.
Look at this.
It even works.
Service detail, a small plumbing repair, helper dashboard shell.
And so, this UI doesn't look too bad.
But there's something that you need to know when you design UI with Codex.
They all look pretty much the same or at least very, very similar.
And that is because Codex is not really an expert when it comes to designing pages.
It's great for coding, for solving all kinds of problems.
But for actually designing stuff, it's not the most ideal tool.
There are a lot of other tools that do it a lot better.
And so, during this specific app build, I want to bring in another tool and show you how we can have a more specialized tool actually design the UI and then bring that UI back into Codex and get Codex to kind of implement it. okay?
Now, when it comes to UI designs, there's lots of cool tools.
There's Google Stitch, there's Figma, and there's also a tool that I really like to use, and it is called FlutterFlow Designer.
It's a very very lightweight tool, and it works really great with Codex.
And so, if you head over to designer.flutterflow.io, you're going to be on this page here.
And you can use the web app version here, but what I really like to do is I like to use their desktop app, okay?
Here's the desktop app, and what's nice about the desktop app is that you have this app integrations.
So, if you click over here, you can connect designer to external agents and IDEs.
So, Claude Code, Gemini CLI, and Codex.
And so, I have this set up.
So, all you need to do is click here.
Now it's installed.
And now, this app here is going to serve as an MCP server, okay?
Because if I go to Codex here, and I press command comma, and I go into MCP servers, I also have this FlutterFlow Designer configured here.
And if you want to set this up, you can simply ask Codex to do it.
It's going to research, it's going to figure out, and it's going to eventually set it up for you so that you can have this FlutterFlow Designer as one of your MCP servers here, okay?
So, now that we have this set up, what we want to do is we want to go into this designer, and we want to click a new design, and here you can enter a prompt ask it to design any kind of app UI for you, whether it is a mobile app or desktop app.
They also have, you know, explore styles feature, or you can have an instant generation.
Now, what you don't want to do is you do not want to type this prompt directly.
You want to ask Codex to generate a prompt for you since it knows a lot more about the app that you can input here.
And this is the same workflow regardless of whether you're using this tool, Google Stage, or pretty much any other tool out there.
And so, we're going to go ahead and we're going to grab a prompt to do just that.
And here's a prompt that we can use.
Create a polished FlutterFlow designer concept for FlexHelp.
FlexHelp is a two-sided marketplace, and we are telling it what kind of screens to design.
We give it some design requirements.
We give give it some guardrails, stuff like that.
And so, we're going to input this prompt, and here we're going to stay on the explore styles, and that way we can have different styles show up, and we can decide which one we like better.
So, I'm going to go ahead and run this, and here are the designs.
So, as you can see, they're all very different.
We have this, we have this, we have this, we have these ones, so and you got to admit that in my view, these look a lot better than what Codex did, right?
So, this is what Codex did.
In my view, these designs are a lot more interesting than Codex's design.
Okay?
So, I don't know which one you guys like.
This is a classical one.
This is nice.
Let's take a look.
Let's go ahead and use this classical one.
So, I'm going to say use this style, and it's going to use this style, and then it's going to generate all the pages that we need.
So, here we have seven pages that we need.
So, now we have all the pages.
Here we have customer home, we have service details, really nice page, booking request form, customer booking status, we even have a map, helper dashboard, helper request management, and a user profile.
The next thing that we want to do is we want to tell Codex to fetch these designs and implement them, bake them into our app.
And so, let's take a look.
So, this is called FlexHelp Marketplace, the name of the project here.
And so, now what we're going to do is we're going to go back to Codex.
We're going to close out this section here, and we're going to give it a prompt that's going to ask it to connect to FlutterFlow designer, look at that specific project and grab those designs and then bake them into our app
right here.
And I have a prompt here to do just that.
So, here's the prompt.
Update the Flex sub static UI to match the approved FlutterFlow Designer reference.
So, this is the main prompt, but we also need to ensure that we tell it that we want to use the FlutterFlow Designer MCP.
Use FlutterFlow Designer MCP and then project is because there are multiple projects.
We want to copy this project here.
This.
And we're going to go into plan mode just to make sure that everything is good.
And I'm going to run that.
And here it says, "Next, I'm opening a Designer MCP session." And once it does that, you can see you can actually see the agent working.
This is our agent, our client cuz Codex is acting as an MCP client connecting to this MCP server.
And as you can see, it's it's analyzing this page.
And you can see this little Codex icon here.
And you can see it says here, "One active agent." It's studying the designs and then it's going to pull the designs back here and build an actual UI using those designs.
And the goal here is to ensure that it builds a pixel-perfect or at least as close as possible to what we see here.
Okay?
So, let's see what it's doing.
So, now it came back with a plan.
Update the existing React static UI to visually follow the approved FlutterFlow Designer reference.
So, here it's it has five pages, but over here we actually have more.
We have seven pages.
So, I don't know if I if it missed it or that I did not it was it wasn't part of the prompt or some mix-up.
So, we want to make sure that it captures all the seven pages.
Ensure that it captures all the seven pages from FlutterFlow Designer project.
Okay?
And still in the plan mode just to make sure we get this right because our app ideally should should have all of these pages.
User profile, helper, request management, helper dashboard, etc. etc.
All of these make sense.
Okay, so it went ahead and it updated it and it's going to add static page components for all seven pages.
Okay, submit this.
All right, so it has implemented the seven page designer reference update.
Let's go ahead and see how it looks.
All right, so this looks a lot better in my opinion.
All right, we have settings.
We do have the CH customer home.
Do we even see that?
I don't even see that here.
But this customer home service details, click on book.
So this menu we need to kind of remove this.
So let me go ahead and sell it.
What are those abbreviations?
Let's see.
Do they need to be there?
Let's see what it says.
But it looks a lot cleaner than than the original UI, at least to my eyes.
And now we're doing a little bit of tweaking, clean-up plan.
Okay, they were placeholder text icons used to mimic the nav icon column.
So we need to remove that and we should probably remove any other placeholder icons.
But right now it looks fairly clean.
Static preview, what is that?
This this also needs to be removed.
So we're going to actually tell it check for any placeholders and see if you can match what it should be on the design.
Pull pull the design elements and re-replace these placeholders with with the actual design.
All right, so it has fixed it.
Now it looks a lot better.
We still see some other static preview.
That's okay.
Let's see if they go away once we have real data being fed.
So once we're happy with the UI, we can start designing the backend.
Data model should match the screens that we actually have here.
And so for FlexHelp, the first schema to support service browsing, helper profiles, booking creation, and helper status updates.
Anything beyond that can wait.
And here's a prompt that we can use to connect our app to Convex, okay?
We're going to be using Convex for this app as well.
So, we have here connect let's help UI to live Convex service and helper data.
Let's go ahead and keep it in plan mode.
All right, so this is our plan.
This is the Convex read path plan.
Connect the customer facing marketplace UI to Convex for the first database back read path.
The repo currently has no Convex setup.
So, add Convex from from scratch, etc., etc.
And we have Convex here running.
We have these three projects.
One that was created for the other app that we did, and so we're going to be creating another project for this app as well.
All right, so it has implemented the first Convex back read path for the customer marketplace.
And these are the files that has changed.
This is what's connected.
And if we jump into our Convex dashboard, we can see we have a brand new project here that wasn't there before.
And if we jump here, we should see that we have a deployment here on my own machine.
This is a local deployment that Convex is running on, right?
This is not on Convex, it's on my own machine.
And we also have a production, right?
So, with this deployment here, you go into data, we can see that we have categories, we have helper profiles.
We also have services.
And then we have a bunch of functions, right?
Marketplace list, and then we have seed, creating seed data.
Now, if you go into production, we may not see anything there, right?
There's nothing there because we haven't deployed this app to production.
There's no need for a production deployment.
Now, if we jump back to Codex and we refresh the app, this is now data that it's pulling from my local Convex deployment, okay?
So, we have here Marcus.
If you go back here, we should see Marcus, right?
Helper profiles, there's Marcus.
And if I change Marcus, if I edit name and I change, let's say, Marcus, right?
And I go back here, now it's automatically Marcus E here.
Automatically updates.
If I change that to James, guess what?
Now it says James S.
So, as you can see, this is now linked to our local Convex instance on my machine.
Both of these are on my machine.
Now, the next thing that I want to do is I want to build out the booking flow, because that's the key in this app, right?
The ability to book these services.
And I have a prompt here as well.
We're going to be using this prompt right here.
Okay, add the customer booking flow to FlexHelp.
Inspect all of this, and I'm going to set this to plan mode.
And let's go ahead and run that.
That's our plan.
We're going to be adding the first customer-side booking creation path on top of the existing Convex back service listing.
And there's some information there.
We're going to implement this plan.
All right, so it has implemented the customer booking creation path.
And it also tested it out.
As you can see, there's also a booking here as well.
So, let's go ahead and see if this works.
All right, so let's say I want to book James.
Click over here.
So, I can't quite pick anything here, but I can do request to book.
Send booking request.
And I need an address.
Okay, so let's say do, I don't know.
We need auto-complete here as well.
Fill in every field.
Okay.
Okay, so task, task.
Okay, this I cannot fill every field.
Okay, customer, schedule, say 10:00 a.m.
And this obviously needs to be, you know, like a calendar picker, that kind of thing.
Okay, and now we have the booking request.
And I can click on view bookings, and I see my bookings.
Okay, so this was done by the Codex while it was testing.
So, now if we jump into Codex, we should see on my local instance here, we should see bookings.
And we see two bookings, okay?
Right here.
So, the booking flow seems to work.
There are some rough edges here and there that we need to fix, but for the most part, it is coming together.
Now, once customers can create bookings, helpers need a place to see and manage them.
The helper dashboard should show incoming requests and make it obvious what action the helper can take next.
At the simplest level, the helper should be able to accept or decline a request.
After accepting, the helper should be able to move the job into progress and then mark it completed.
Those status changes are what turn the app into a workflow.
This is also a good place to show the power of live data.
When a status changes, the UI should update and the Convex row should change.
That gives us a real end-to-end marketplace loop.
Customer request, helper action, back end state, and UI update.
And this is the prompt that we're going to be using.
Add the helper dashboard workflow for viewing booking requests and updating booking status.
Okay, I'm going to switch into plan mode and we're going to go ahead and send this.
So, it's asking me that how should the helper workflow handle the existing requested booking status?
Keep requested, this is recommended.
Preserve existing created bookings, show it as pending in the UI, or rename to pending.
Change the stored initial status to pending and require a small data.
So, we don't want any migration.
We just want to go with the easiest option.
And then, how should the helper request visibility work before auth is added?
Helper selected.
This is add a simple static helper switcher filter, so each helper sees bookings for their Convex helper profile only.
Simplest UI and back end, but less realistic.
So, we want helper selected.
Okay, we're going to select that.
All right, there's the plan.
Add the first helper-facing workflow on top of existing convex booking creation pack.
We're going to submit this.
Here is our progress.
All right, so he just implemented the helper booking status workflow.
And these are the files that changed.
This is what it did.
So it updated it through pending, accepted, in progress, completed.
It viewed it under helper James created booking this.
So it did all that.
It did some tests.
Let's go ahead and open it.
And let's see what we can do.
So this is the helper booking status.
I'm going to helper dashboard have my incoming requests, right?
So I can accept it, I can start it, and I can complete it.
Okay?
So that's the piece that is working.
I can also cancel it.
And that is canceled.
And that's it.
I can't come back to it.
I can't do anything.
Here we have recent activity, right?
Completed, completed, canceled.
And this is the one that did 148 116.
This is the one that it has completed.
This seems to work, right?
And we have canceled, we have completed.
Even a simple marketplace has different responsibilities.
Customers create booking requests and view their own bookings.
Helpers manage requests that belong to them or are available to them.
Admin or system logic may handle category alteration or platform level controls later.
The key point is that these rules should not only live in the UI.
Back-end functions need to enforce ownership and permissions because the UI can always be bypassed.
This is where Codex needs to think carefully instead of just adding buttons.
And so now what we want to do is we want to add the roles off and the access boundary layer.
And here's the prompt to do that.
Add the first practical roles off and access boundary layer around the existing flex help workflows, okay?
So we're going to go definitely in the plan mode.
We're going to run this and see what it says.
And so it has a question for us.
Which identity approach should this plan should this step plan for?
Demo sessions.
Add Convex back demo users and for Convex or add real Convex or sign and sign up more production shape but larger and currently beta or this clerk use third-party or provider.
Let's go with Convex or how should customer helper roles be represented for or version one?
Role plus profile.
So what is this?
Each demo user has customer helper both and helpers linked to one helper profile for request.
Separate accounts.
Customer and helper are always separate demo users.
Simpler boundaries but less flexible.
And sometimes when I'm not sure but I want the system to come up with a really solid answer.
This is what I usually type.
For the second question, what's the most durable long-term option?
Okay, that's the key here, durable.
Now let's see what it says.
The most durable long-term option is role plus profile.
Perfect.
That's it.
And for the first question, choose demo sessions for the for this core step.
Okay, so let's go ahead with demo sessions, okay?
So do demo sessions and role plus profile.
Let's just go with the recommendations.
Expand it.
All right, so now you can see that Codex is testing everything, right?
It's testing on the UI in the browser and you see this cursor right here.
The browser flow now confirms a customer created book booking appears for the linked helper and can be moved into in progress.
All right, so it has finished and it has implemented the demo roles and access boundary layer over here, okay?
So marketplace browsing stays public.
Booking creation requires a customer or both demo user.
Customer booking status only shows bookings owned by the signed-in customer.
Helper request status updates require a helper or both demo user with a linked helper ID helper profile ID.
And so, remaining limitation, this is still tutorial demo auth, not production authentication.
Real Convex auth or Clerk can replace demo session issuance later while keeping the same role profile access model.
And so, this should give us some uh role separation, right?
Some kind of authentication on the backend.
And here, if we jump into Convex, you can see we now have users over here, right?
These are the users, and there's a role.
So, you can be either both, or you can be a helper, or you can be a customer.
And depending on which role you're you're in, you can do various operations.
Here we have sessions, we have various services, we also have functions that also handle these roles.
So, you can list demo users, you can start demo session, viewer list for helper, list for customer, etc., etc.
And these functions, they are meant to work with specific roles, right?
So, that authentic- -ation is is baked in into these functions.
Now, when something breaks, the wrong move is to immediately ask Codex to rewrite a whole section.
The better move is to give it the error, ask it to inspect the relevant files, and make it explain the likely cause before changing anything.
That debugging loop is the same whether the problem is a browser error, a build failure, a Convex function error, or a UI state that does not make sense.
Observe the issue, diagnose it, patch it narrowly, and verify the fix.
And so, let's give it a prompt that will deal with this sort of issue regarding debugging, polish, verification, stuff like that.
We're going to be using this prompt over here.
Debug, polish, and verify the existing flex help product workflow.
Inspect the current app first, then identify this is smallest set of fixes and polish improvements needed to make the current workflow feel credible.
Let's jump into plan mode.
Right away, it tells us that I see a few concrete polish fix candidates already and it lists some here, okay?
And so, the build passes and both local servers are responding.
So, it's using an in-app browser now for a quick workflow probe.
So, the plan targets actual UI issues rather than guesses, okay?
So, let's wait to see what it finds.
And this is the plan that it came up with.
Tighten the existing customer helper Convex and demo app workflows without adding new product scope, okay?
So, it lists a bunch of things and we want to quickly scan assumptions.
No new marketplace features will be added, no payments, nothing, okay?
That stays and here's some key changes.
You can kind of scan through that.
Let's go ahead and submit this.
Right, so it is done and it has implemented the debug polish task.
And it actually fixed a lot of interesting things.
It fixed some placeholders, it did some testing, it switched to different accounts, actually cleaned up a lot of cool stuff.
So, if we take a look here, we can see that we can book, we can request a book, we can describe the task, we can fill all of this out, send, add the preferred date, and now it's showing in that I'm signed in as the customer, right?
And I can switch a demo user.
Now, I'm signed in as the helper and now I'm signed in as both.
Customer access, choose a customer demo user to create a booking request.
So, it's coming along really, really nicely.
And the final build step is deployment.
The app is not done until the front end is live, connected to the intended Convex back end, and smoke tested through the real product workflow.
So, let's go ahead and grab this prompt and let's go ahead and paste it.
So, deploy the existing Flex app with Vercel and smoke test the connected Convex backend, okay?
We're going to jump into plan mode so that it can prepare everything, plan everything out.
One thing that I want to tell it that I didn't tell it, and I'm going to do that right now, is I want to make sure it uses the Vercel package.
Use the Vercel Vercel plugin.
And I'm going to press command enter so that it steers the conversation because we want the model to ensure that it catches that, right?
That we want to use the Vercel plugin.
Which Convex environment should the deployed Vercel app use for the smoke test?
We want to use production.
And where should the front end be deployed?
New personal project.
Exactly.
This is a brand new project.
We already have a bunch of projects, right?
So, if we go over here, we have this project.
Actually, we should have used the existing project.
We already have this project, but I think it's going to create a new project.
Let's see if we can tell it.
I'm going to copy that.
See if we can use the existing project.
No need to create a new project.
And let me check.
So, this is actually plan mode, so we're okay.
Okay, implement this.
And here it's going to create a production.
So, if we go over here, remember we're using this.
We also have a production deployment set up.
So, it's going to be using the production deployment.
So, we should see data and tables and functions and all that good stuff as part of this deployment.
Production Convex is now seeded and responding with services and demo users.
So, if you go over here, we have everything now.
Look at this.
Sessions, services.
We have the data.
All right, so looks like it has deployed it, and now it's doing a couple of smoke tests to make sure that everything works in production because now it's in production and it's also using the production Convex deployment, right?
So, we need to test things again because, you know, it's using new environments now.
All right, and it is done.
Deployed FlexHelp.
This is the URL.
This is the Vercel project deployment ID.
Everything is live now, okay?
And you can access it as well at this URL, okay?
So, we have customer home, and I'm signed in as one of these demo users, customer or helper, right?
I can go into helper dashboard, and let's say I go I sign in as a customer, and I go into customer home, but I can't click here, right?
This is helper access, right?
Remember the roles, right?
It's it's being handled here.
And so, now as a customer, I want to book professional seasonal garden cleanup.
I'm going to book this, request to book, I'm going to fill all of this out, 10:00 a.m., 2 hours, it's fixed address, send booking.
Okay, so with booking request submitted, your request was saved, view my bookings.
So, we have all of these pending, and so now I'm going to log in as Marcus, and Marcus is going to go into helper dashboard, and now he sees the booking.
And I'm going to say accept or complete, it's completed.
Now, if I come back as Anders again, and I go into customer's home, booking request for maybe summary, right here, customer booking status.
Okay, there we go, completed.
And here we see completed.
This one is completed.
Actually, this one, for Marcus, it has been completed.
And then there's some more requests here.
So, the flow is working.
The flow is working perfectly.
Of course, there are so many other things that we can do, we can use real auth, we can change the page, we can update things, but right now, the core of the app works perfectly.
And just as a little bonus, when you reach this stage, there's more things that you can do, but one thing that, you know, right away that you can ask Codex to do would be something like suggest top five high impact, low effort features,
functionalities.
And I'm going to switch into plan mode, and we want to see if there is some high amp impact features that we can add that do not require a lot of effort.
And here they are, top five high impact, low effort next features I prioritize.
Functional marketplace search, booking detail view, customer cancel pending requests, etc., etc.
And if you don't want to implement these right now, what I like to do is save this list as dots.todo.md as a markdown file and prioritize each item on a scale of one to five, one being urgent.
And so that way I have a constantly evolving list of things that I can build in the future.
So I do not need to ask Codex all the time constantly, you know, you know, to to suggest new features.
And I'm going to say submit and that's going to go create this dots directory as well and this todo.md, which is the file that I usually keep to store new features.
And now we can open this file and say, "Codex, so this is priority one, two, three, four, five." So now we have a an app that is live, that is on the internet, that we have built, that works.
This workflow, this main workflow works as expected, but of course, there's so many things that you can do to improve this app.
And in some of the later videos, I'm going to be showing you some more advanced Codex techniques that you can use to make your apps even better and more robust because I've been using Codex for a while now and I have some advanced tips and tricks that I'd
like to share you in some of the future videos.
And so I really hope that you enjoyed this mega Codex tutorial.
Now, if you want to go ahead and build this app or modify this app or whatnot, I'm going to be leaving all my prompts, every single prompt that I used to build this app inside of our amazing Patreon community and you can check it out and
hopefully join.
There are lots of great people there and of course, support this channel, and support my work.
And so, like this video, leave a comment, and check out our amazing Patreon community.