0:00.000–0:01.597
If you're watching this video, then
0:01.597–0:03.140
I'm going to guess that you're some
0:03.140–0:06.280
body who's been running into plan limits on your Claude subscription.
0:06.880–0:10.305
Well, the good news for you is that I'm about to show you 11 ways so
0:10.305–0:13.600
you can make sure you never run into those usage limits again.
0:13.860–0:17.669
And there's a key fundamental principle that most people totally overlook,
0:17.669–0:21.120
and it will change the way that you interact with Claude code forever.
0:21.760–0:27.260
So the first thing which you need to understand is that Claude code does not actually remember anything at all.
0:27.260–0:31.844
So the way this actually works is for every new message you send away to CloudCode,
0:31.844–0:36.631
it's not only receiving that latest message, it's also receiving everything prior t
0:36.631–0:37.440
o that as well.
0:37.900–0:40.358
So if you've had 50 interactions back and
0:40.358–0:42.888
forth with CloudCode in your session, the
0:42.888–0:46.720
n on the 51st message, it's receiving all of that context again.
0:47.220–0:48.813
And you might be thinking, well,
0:48.813–0:51.940
that would be chewing through millions of tokens, and it would.
0:51.940–0:54.300
And that brings us on to the idea of caching.
0:54.300–1:01.600
So caching is how Claude Code knows what information already existed in the session versus what is new.
1:02.020–1:07.680
And this is what stops us from burning through our session limit after maybe only 20 or 30 minutes.
1:08.260–1:12.011
So on the first turn, so your first interaction in that session,
1:12.011–1:15.832
Claude is going to receive the system prompt, the project contex
1:15.832–1:17.460
t and your initial message.
1:17.780–1:18.868
Then on turn two,
1:18.868–1:20.035
Claude knows that
1:20.035–1:24.000
all of this information from the first turn already existed.
1:24.300–1:28.811
So all you're going to be charged for on turn two is Claude's response and
1:28.811–1:30.240
then your next prompt.
1:30.700–1:31.960
And it's the same for turn three.
1:32.320–1:35.840
It then knows that all of the information from turn two was already there.
1:36.200–1:39.360
So you're only being billed for the new stuff at the end.
1:39.360–1:42.414
And everything which happens before this reply,
1:42.414–1:44.500
all of this is known as the prefix.
1:44.840–1:47.887
You'll see in a second why this is so important when
1:47.887–1:49.120
we get into the tips.
1:49.120–1:54.800
but for now just understand that if something in the cash data or something in the prefix changes
1:54.800–2:01.060
then you're going to be paying for all of this a second time which means if you have 500 or 600
2:01.060–2:03.785
thousand tokens sitting in your context window and
2:03.785–2:06.320
something in the prefix has changed then you're
2:06.320–2:08.411
going to be paying for all of that again. So
2:08.411–2:10.860
all of that is to say that Claude does not remember
2:10.860–2:12.498
who you are between turns.
2:12.498–2:16.520
Every new message it's sending everything back to Claude again,
2:16.520–2:19.613
it's going to read through it, it knows what was already there,
2:19.613–2:21.160
it knows what is new, and you're
2:21.160–2:24.498
only going to be billed full price for the new stuff.
2:24.498–2:27.060
Now if you're using it on the API, we know
2:27.060–2:31.468
that cached inputs are 0.1x of the initial input cost.
2:31.468–2:34.700
CLODs do not disclose publicly in their
2:34.700–2:38.773
documentation what that multiple is for planned subscriptions,
2:38.773–2:40.920
but we do know that we're not paying
2:40.920–2:42.694
full price for those tokens.
2:42.694–2:44.320
Again it's going to be some
2:44.320–2:46.980
form of fraction of the initial input cost.
2:47.100–2:51.500
Now that we've cleared up that important distinction everything from here on will make much more sense
2:51.500–2:56.920
so let's move into the first tip which is keeping all of your important context inside of your
2:56.920–2:58.439
clod.md file. A clod.
2:58.439–2:59.875
md is just a markdown
2:59.875–3:03.420
file which lives inside of your project and it gets
3:03.420–3:09.460
auto-injected into your context window at every new session start in that project. So the reason
3:09.460–3:12.193
this is so important is if you're not using a clod.
3:12.193–3:14.660
md file then at the start of every new session
3:14.660–3:17.967
you need to explain who you are, what your tech stack is,
3:17.967–3:20.340
where your files go, what your rules are
3:20.340–3:23.955
and then you need to give it your task or your initial prompt.
3:23.955–3:25.980
Whereas if you are using a clod.md
3:25.980–3:28.636
all of this context is already stored in there so
3:28.636–3:30.960
clod already has all of that information.
3:31.520–3:33.492
You're not wasting tokens going back and
3:33.492–3:36.160
forth explaining the same thing every time. So it won't
3:36.160–3:38.820
just save you tokens, it will also save you time.
3:38.820–3:41.480
You can see my clod.md inside of my project here.
3:41.880–3:43.482
Now if you want to create one,
3:43.482–3:46.620
you can run slash command slash init and you can see here,
3:47.060–3:48.383
initialize a new clod.
3:48.383–3:49.706
md file with code-base
3:49.706–3:52.560
d documentation or you can just describe in plain
3:52.560–3:55.307
English that you want it to create you a clod.md and
3:55.307–3:57.400
it will go ahead and do that for you. Tip
3:57.400–4:01.268
number two is keeping your context below 20 to 25%.
4:01.268–4:04.400
If Opus 4.8 and these other new anthropic
4:04.400–4:07.486
models all have a million tokens of context then
4:07.486–4:10.340
why would we aim to keep things sub 25 percent
4:10.340–4:13.104
well there's a multitude of different reasons for this but
4:13.104–4:14.740
i'll just cover a couple of them now
4:14.740–4:20.280
and the next three tips after this one are all related back to this same issue so this is very
4:20.280–4:24.295
very important so the first one is as i mentioned in the intro when
4:24.295–4:26.080
we send our prompts the first
4:26.080–4:31.880
time around to the clod api in clod code we're being charged 100 of the cost for those tokens
4:31.880–4:35.260
So we're being billed one token for every token.
4:35.260–4:37.880
Now after that, once those tokens are
4:37.880–4:40.029
cashed and they become part of the prefix,
4:40.029–4:42.240
we're still paying for those tokens. And we
4:42.240–4:47.160
don't know what that cost is for Anthropic subscriptions or Claude plans, but we do know
4:47.160–4:51.192
that those tokens are not free. So even if it's 0.
4:51.192–4:54.820
05x of the initial cost, so 1 20th of a token,
4:55.240–4:58.557
if you have 500,000 tokens in your context window,
4:58.557–5:01.400
then you're paying 25,000 tokens per turn,
5:01.400–5:05.864
and that's going to be increasing with every new turn just in cached reads.
5:05.864–5:07.160
So that's one reason.
5:07.600–5:13.000
Now another reason that you would want to keep your context sub 20% is we know that these models
5:13.000–5:17.378
start degrading in performance the more you start to fill this context up.
5:17.378–5:18.720
So what that means is if
5:18.720–5:22.588
you have 800,000 tokens, so 80% of your context window used,
5:22.588–5:24.600
then the model is not going to be
5:24.600–5:28.040
performing optimally as it would if we only had 150,
5:28.040–5:30.840
000 tokens in that context window. You can
5:30.840–5:35.030
see down here in my status line that I have my context usage as a bar and
5:35.030–5:36.280
a percentage listed
5:36.280–5:39.059
there which means I can always see it at a glance.
5:39.059–5:41.700
Now I would recommend that you do the same and I
5:41.700–5:44.207
will put a prompt on screen which you can copy and
5:44.207–5:46.400
give to Cloud Code so that it can implement
5:46.400–5:47.754
the same thing for you.
5:47.754–5:48.966
If you want to keep an
5:48.966–5:51.960
eye on your context just run forward slash context
5:51.960–5:55.095
and it will bring up this visual where you can see what you've used.
5:55.095–5:56.520
You can see it broken down here
5:56.520–5:59.251
on the right hand side as well and it breaks it down below.
5:59.251–6:00.820
When you start approaching that
6:00.820–6:05.886
20 to 25% limit, I recommend that you ask Claude to generate a handoff file,
6:05.886–6:07.520
which is just a markdown
6:07.520–6:11.095
file with a summary of everything you've done in that session,
6:11.095–6:13.020
and also give you a session resume
6:13.020–6:14.395
prompt. So you can then
6:14.395–6:15.841
give that session resum
6:15.841–6:18.880
e prompt back to Claude in the new session, or after
6:18.880–6:22.062
you've cleared your context, have it read the handoff file,
6:22.062–6:23.780
and you can pick up exactly where
6:23.780–6:26.913
you left off. And if you want to go one step further,
6:26.913–6:29.300
I built this into a skill, I just called
6:29.300–6:32.971
it handoff so all I need to do is run forward slash handoff and
6:32.971–6:34.660
it will do all of that for me.
6:34.920–6:39.560
If you want access to the skill you will find all of the resources from the video in the pinned
6:39.560–6:40.667
comment down below.
6:40.667–6:41.709
Tip number three is
6:41.709–6:44.640
making sure that you keep your prompts within one hour
6:44.640–6:47.492
of the previous prompt. This is one of the simplest and
6:47.492–6:49.900
most common mistakes that I see people making
6:49.900–6:51.365
inside of Claude Code and
6:51.365–6:52.829
it will chew through your
6:52.829–6:55.480
five hour session limit and your weekly quota
6:55.480–6:57.237
on your Claude subscription. So
6:57.237–7:00.880
we're looking at the Claude Code documentation here and you can see
7:00.880–7:04.851
that it says Claude Code requests the one hour TTL automatically.
7:04.851–7:06.440
So we're talking about the
7:06.440–7:09.440
caching again here. So when you use Claude Code,
7:09.440–7:12.140
every time you send a new message away, that
7:12.140–7:16.500
resets the one hour cache. Now if you go out with that one hour,
7:16.500–7:18.980
then the cache is then invalidated
7:18.980–7:20.094
and your next prompt,
7:20.094–7:21.145
you're going to be pa
7:21.145–7:23.620
ying for all of those tokens again. So if you have
7:23.620–7:25.893
500,000 tokens sitting in your context window,
7:25.893–7:28.280
which you shouldn't have if you're following step
7:28.280–7:31.543
two, then if you go away for an hour and you come back,
7:31.543–7:33.820
you send your next prompt to Claude,
7:34.160–7:37.772
you're paying full price for those 500,000 tokens again.
7:37.772–7:40.180
If you then go away for another hour and
7:40.180–7:42.530
come back and then send your next prompt,
7:42.530–7:45.640
you've now paid for those 500,000 tokens three times.
7:46.040–7:49.980
So this is why it's super important that you keep it within an hour of your previous prompt
7:49.980–7:54.380
and going back to the previous tip that you keep your contacts below 20%
7:54.380–7:57.930
so if you do go out with that hour you're not burning 500,
7:57.930–7:59.860
000 tokens every single time.
8:02.020–8:03.106
model mid-session.
8:03.106–8:04.063
How many times hav
8:04.063–8:07.320
e you heard that you should be changing your model from Sonnet
8:07.320–8:09.708
to Opus for more complex tasks and then
8:09.708–8:13.140
back to Sonnet again for the easier stuff? And I'm going
8:13.140–8:15.243
to tell you now that that is bad advice.
8:15.243–8:18.200
We're staying inside of the Cloud Code docs and you can
8:18.200–8:21.863
see here actions that invalidate the cache.
8:21.863–8:26.020
Number one on the list is switching models. So if we
8:26.020–8:29.621
scroll down you can see that it says each model has its own cache.
8:29.621–8:31.320
Switching with forward slash
8:31.320–8:38.040
model means the next request reads the entire conversation history with no cache hits even
8:38.040–8:39.895
know the content is identical. So
8:39.895–8:43.540
I see people making this mistake all the time. They're constantly
8:43.540–8:47.020
switching models between Opus, Sonnet, Sonnet, Opus,
8:47.020–8:49.820
maybe Fable, and you're paying for all those
8:49.820–8:52.198
tokens again every single time you make that switch.
8:52.198–8:53.820
I'm not saying that you should not be
8:53.820–8:55.386
switching models at all because
8:55.386–8:58.460
that is genuinely good advice to use the right model based on the
8:58.460–9:04.700
task at hand. What I am saying is that you should not be switching model mid-session. You should be
9:04.700–9:07.101
putting together a plan using your higher power model,
9:07.101–9:08.980
whether that's Opus or whether it's Fable,
9:09.440–9:11.753
and then once you have a structured plan in place,
9:11.753–9:13.840
switch your model down to whatever it is that
9:13.840–9:16.160
you're going to be using to execute that task, and
9:16.160–9:18.140
do not switch again unless you absolutely
9:18.140–9:21.084
have to. Number five is almost identical, and
9:21.084–9:23.640
that is avoiding changing effort level
9:23.640–9:26.657
mid-session. Right below switching models and
9:26.657–9:29.900
actions that invalidate the cash, we have changing
9:29.900–9:30.934
effort level. So
9:30.934–9:31.894
you can see here
9:31.894–9:35.440
that it says that the cache is keyed by effort level as well
9:35.440–9:41.120
as model so switching with forward slash effort means that the next request reads the entire
9:41.120–9:43.660
history again with no cache hits. So
9:43.660–9:47.640
this is the exact same as the previous tip if you can avoid
9:47.640–9:49.816
changing effort level mid-session then
9:49.816–9:53.080
you should absolutely be doing that. Number six is building a
9:53.080–9:54.273
proper memory system.
9:54.273–9:55.277
I have a full video o
9:55.277–9:58.040
n my channel breaking down how you can set this system
9:58.040–10:01.190
up for yourself. I'm not going to go too deep here but
10:01.190–10:03.680
I will link that video on screen. But what
10:03.680–10:07.632
we're talking about doing here is building a proper memory system so
10:07.632–10:08.880
that Claude Code has a
10:08.880–10:10.551
working map of our project and
10:10.551–10:13.960
it always knows where to go to find the information that we're
10:13.960–10:17.101
asking for. And we can do that by using something like Obsidian.
10:17.101–10:18.820
And if you don't know what Obsidian
10:18.820–10:21.753
is, it's just a software that allows us to view and
10:21.753–10:24.400
edit marked-in files. And you can see here an
10:24.400–10:29.280
example of my obsidian knowledge graph you get this really cool graph we can kind of like pull
10:29.280–10:33.320
around all these different nodes all these different lines that you see are connections
10:33.320–10:35.127
between different documents so
10:35.127–10:38.340
it can piece together different bits of information work
10:38.340–10:42.480
between different documents to find the right information to give you a high level overview of
10:42.480–10:46.248
how this works we create a folder somewhere on our computer and
10:46.248–10:48.060
we then point obsidian at this
10:48.060–10:49.677
folder and then that then
10:49.677–10:51.294
becomes our vault now onc
10:51.294–10:54.220
e we have our vault we can then point cloud code
10:54.220–10:55.305
at this vault and
10:55.305–10:56.390
inside of our vau
10:56.390–11:00.420
lt we will have a bunch of different folders in there. Now these
11:00.420–11:02.473
folders can be whatever you like.
11:02.473–11:06.140
For me I like to split this into three separate layers. So the
11:06.140–11:08.627
first layer is my working folders. So
11:08.627–11:12.880
for me that's things like agency, school, content and inside of
11:12.880–11:15.167
these folders I will have subfolders. So
11:15.167–11:18.260
for agency that will be things like clients, it could be
11:18.260–11:21.464
deliverables, outreach, pipeline, for school,
11:21.464–11:24.980
it might be frameworks, strategy, courses. So that is
11:24.980–11:25.923
our first layer,
11:25.923–11:26.799
all of the worki
11:26.799–11:30.100
ng files that you'll be in there messing around with on a day
11:30.100–11:35.640
to day basis. The second layer here originates from Andre Carpathie's Obsidian RAG system. I'm
11:35.640–11:37.077
not going to go into that here so
11:37.077–11:39.840
make sure you check out that video linked on screen if you're
11:39.840–11:43.045
interested in that. And the third layer is our clod.
11:43.045–11:45.580
md. So in there as we covered in tip number
11:45.580–11:48.790
one. This is where we store our project map, our rules,
11:48.790–11:51.500
our context, so that Claude knows everything
11:51.500–11:56.200
which it needs to know about us at session start without us having to explain every time. So how
11:56.200–11:59.454
that relates to our memory system here is, first of all,
11:59.454–12:02.000
we give Claude a map of the project folders,
12:02.440–12:04.609
what's contained in each folder, so
12:04.609–12:08.080
that it knows where things are. Then we need to give it a map
12:08.080–12:09.982
of where it needs to put things when
12:09.982–12:13.000
we ask for them to be saved. So you can see here, anything
12:13.000–12:15.814
for a client goes in agency clients and then
12:15.814–12:18.940
the client's name. If it's an idea or a script or a
12:18.940–12:22.261
hook it goes in content. So we give it the information which it needs so
12:22.261–12:23.120
that things don't
12:23.120–12:25.227
get lost when it saves them. And
12:25.227–12:29.280
finally we give it the knowledge base rules. So this is how it
12:29.280–12:31.953
knows how to operate our knowledge base system and
12:31.953–12:34.180
this is arguably the most important part.
12:34.540–12:38.420
You can see here we have information on wiki links to make sure that it's always creating
12:38.420–12:41.180
those links between different documents and
12:41.180–12:43.940
that is how we get our nice graph here inside of
12:43.940–12:47.550
obsidian.md. Number seven is creating custom skills.
12:47.550–12:50.140
This one is hands down the most powerful
12:50.140–12:53.841
tip in this video, not only from a token saving perspective,
12:53.841–12:55.840
but also in maximizing the value
12:55.840–12:59.265
which you get out of cloud code for you specifically.
12:59.265–13:02.300
So creating a custom skill is just packaging up
13:02.300–13:07.420
some amount of rich information on a specific topic that's specific to you so
13:07.420–13:08.860
that it's repeatable
13:08.860–13:10.393
and Claude can do it over and
13:10.393–13:14.060
over again without us having to go back 20 different times, change
13:14.060–13:15.718
this, change that, because
13:15.718–13:19.900
Claude already knows what a good output looks like. Now I would highly
13:19.900–13:23.878
recommend that you install the skill creator skill by Anthropics.
13:23.878–13:26.080
This is an official Anthropics skill
13:26.080–13:29.038
and it tells Claude Code how to build custom skills.
13:29.038–13:31.240
So first of all you're going to explain
13:31.240–13:33.035
what it is that you're trying to do. Then
13:33.035–13:35.320
Cloud Code is going to ask you a bunch of questions,
13:35.720–13:37.914
it's going to test it, it's going to run evaluations,
13:37.914–13:39.160
it's going to make sure that it
13:39.160–13:40.434
triggers when it should, and
13:40.434–13:43.300
it's going to make sure that you're happy with the output. So then
13:43.300–13:44.497
all you need to do when
13:44.497–13:45.694
you want to use it goin
13:45.694–13:48.420
g forward is run a slash command and then that task
13:48.420–13:52.119
is now repeatable on the mat. So the reason that this is so
13:52.119–13:54.480
powerful is without a skill, you ask
13:54.480–13:55.819
it to do something, you then
13:55.819–13:59.020
have to go back saying no that's not right, you might need to go back
13:59.020–14:00.579
again and again and again and
14:00.579–14:04.280
eventually you will have a product that you're happy with but by that
14:04.280–14:06.408
point you've burned a bunch of tokens and
14:06.408–14:09.100
if you're doing this multiple times a week then I can
14:09.100–14:14.840
guarantee you're wasting a huge amount of your plan limits on tasks that are repeatable. So what I
14:14.840–14:18.473
would suggest is that you copy the prompt which you see on screen and
14:18.473–14:19.900
you have Claude go through
14:19.900–14:21.869
your last 10, 15, 20 sessions and
14:21.869–14:25.660
pick out things which you've asked it to do multiple different
14:25.660–14:28.397
times and tell it to build those into skills.
14:28.397–14:31.060
To install the skill creator skill just run
14:31.060–14:33.063
forward slash plugins and then
14:33.063–14:37.300
on the discover tab here you will see it listed fourth from the top.
14:37.440–14:40.244
You can see skill creator create new skills,
14:40.244–14:43.200
improve existing skills and measure skills as
14:43.200–14:45.796
well. So all you need to do is enter this and
14:45.796–14:48.540
install it. I would recommend that you go for
14:48.540–14:52.540
user scope which means you'll have access to across all of your different projects.
14:52.540–14:54.640
Number eight is using Caveman.
14:54.980–14:58.320
Caveman is a plugin that makes Cloud Code less verbose.
14:58.780–15:01.700
So we know that these large language models love to talk.
15:02.040–15:04.880
You ask it a simple question and it gives you a huge reply,
15:05.200–15:06.820
which will take you five minutes to read.
15:07.440–15:10.600
And Caveman was designed to try and stop that from happening.
15:10.840–15:12.600
And it actually does a very, very good job.
15:12.980–15:14.028
I've been using it on and
15:14.028–15:16.020
off now for around about three or four months.
15:16.280–15:17.700
And it is very effective,
15:17.880–15:21.740
especially if you're not wanting a big walls of text back from the model.
15:22.320–15:26.240
So it says here, why use many token when few do trick?
15:26.580–15:27.700
And that is kind of how it works.
15:28.120–15:30.520
Make your AI coding agent talk like a caveman.
15:30.520–15:34.420
Same answers, up to 65% fewer output tokens.
15:34.520–15:35.923
Now for me personally,
15:35.923–15:39.540
I probably haven't seen as much as 65% fewer output tokens,
15:39.700–15:42.800
but it does definitely reduce the token output.
15:43.020–15:44.100
You can see an example here.
15:44.220–15:48.200
I said, explain how prompt caching reduces cost in LLMs.
15:48.200–15:51.900
and this was the response that I got using Caveman.
15:52.360–15:54.260
Now if we compare that to this other terminal,
15:54.740–15:56.140
I gave it the exact same prompt,
15:56.260–15:59.320
explain how prompt caching reduces cost in LLMs,
15:59.320–16:00.000
and...
16:00.000–16:01.174
reduces cost in LLMs and
16:01.174–16:02.349
the response which I got
16:02.349–16:04.580
from this one was much much longer. You can see
16:04.580–16:07.018
the difference here in the size of the response. So
16:07.018–16:09.340
this is definitely one that you should be trying
16:09.340–16:11.350
out. Number nine is using ponytail.
16:11.350–16:14.700
This one has been blown up over the past few weeks. You can see
16:14.700–16:17.345
that it's number one repository of the day and
16:17.345–16:20.060
of the week and it was designed because these AI
16:20.060–16:23.672
coding agents love writing unnecessary amounts of code and
16:23.672–16:26.200
ponytail is essentially designed to make
16:26.200–16:30.474
your AI coding agent as lazy as possible without making it complacent.
16:30.474–16:31.560
So if we look at the
16:31.560–16:34.817
benchmarks here you can see the impact which this has.
16:34.817–16:37.060
So this is versus a no skill baseline,
16:37.060–16:42.642
minus 54% on lines of code, minus 22% on tokens,
16:42.642–16:48.080
minus 20% on cost, and minus 27% on time. And if
16:48.080–16:49.801
we keep scrolling down here,
16:49.801–16:53.600
you can see this hierarchy of how it works. So before writing any
16:53.600–16:56.335
code, the agent will stop at the first rung that holds.
16:56.335–16:58.340
The first thing which it's going to do is
16:58.340–17:01.605
check, does this even need to exist? Only if that is a yes,
17:01.605–17:03.480
is it then going to see if there's
17:03.480–17:06.468
something already in the code base that can handle it.
17:06.468–17:08.460
And if that is a no, it's going to see,
17:08.900–17:11.781
can I handle this with one single line of code? And
17:11.781–17:13.960
only if that is a no, is it then going to
17:13.960–17:16.850
write the minimum amount of code to get the job done.
17:16.850–17:19.000
Number 10 is using the Codex plugin for
17:19.000–17:23.980
CloudCode. This is an amazing plugin if you're someone who's already paying for an OpenAI
17:23.980–17:29.700
subscription. So it allows us to use Codex inside of the CloudCode ecosystem,
17:29.700–17:31.020
which means we can
17:31.020–17:35.415
offload tasks from the Anthropic models onto the GPT models.
17:35.415–17:38.000
Now we know that these GPT models are
17:38.000–17:40.867
substantially cheaper than the Anthropic models and
17:40.867–17:43.160
they're now very very competitive as well.
17:43.700–17:45.341
GPT 5.6 has just been released so
17:45.341–17:48.380
that is definitely something that you can take advantage of
17:48.380–17:50.795
if you're already paying for an OpenAI plan.
17:50.795–17:53.080
Once you have the plugin installed just do
17:53.080–17:54.187
forward slash codex and
17:54.187–17:57.120
you will see all the different slash commands which you now have
17:57.120–18:01.553
available. One which you should definitely check out is Codex Adversarial Review,
18:01.553–18:02.440
which is going to
18:02.440–18:07.180
have the Codex agent do a deep dive into your code base to see if it can find any security
18:07.180–18:10.405
vulnerabilities. This is definitely one that you should check out.
18:10.405–18:11.820
Go in there, have a play around
18:11.820–18:14.279
with it, and this will save you a bunch of tokens.
18:14.279–18:16.860
And finally, tip number 11 is offloading research
18:16.860–18:20.673
heavy tasks to Google servers using Notebook LMPi.
18:20.673–18:24.220
Notebook LMPi allows us to use Google servers to
18:24.220–18:27.231
carry out any kind of deep web research. We can also
18:27.231–18:29.740
use it to generate infographics and slide
18:29.740–18:32.146
decks and a bunch of other cool stuff as well and
18:32.146–18:34.120
we can do this all from the command line
18:34.120–18:35.611
inside of Cloud Code.
18:35.611–18:36.937
Now one of the main b
18:36.937–18:40.500
enefits of this is that we know that research flows,
18:40.920–18:44.610
sub-agents, agent teams, all these things are super token heavy and
18:44.610–18:45.840
we don't really want to
18:45.840–18:49.428
be chewing through our Cloud subscription on those tasks and
18:49.428–18:51.680
that is when we can offload those tasks
18:51.680–18:55.781
to Google servers essentially for free using Notebook LMPi and
18:55.781–18:57.560
one of the massive benefits
18:57.560–18:58.581
to this is because
18:58.581–19:02.460
it's Gemini under the hood it has access to all of the YouTube content
19:02.460–19:03.304
transcripts and
19:03.304–19:04.027
all of that ric
19:04.027–19:07.100
h information which you normally would not be able to surface
19:07.100–19:09.317
just using Claude Code alone. So
19:09.317–19:13.340
I asked Claude Code to use the Notebook LMPi CLI to carry out
19:13.340–19:17.043
research on GPT 5.6 SOL versus Claude Fable 5 and
19:17.043–19:20.560
then turn that into an infographic. So it went
19:20.560–19:23.516
through the process here and this is what we got at the end.
19:23.516–19:25.240
Now bear in mind I did not need to do
19:25.240–19:29.880
anything for this. I did not use any of my clod tokens or my plan limits and
19:29.880–19:31.040
this is what we got
19:31.040–19:33.252
at the end. So this is an awesome tool and
19:33.252–19:36.000
you should definitely be leveraging this to save
19:36.000–19:39.481
yourself a bunch of tokens. That is everything for this video today.
19:39.481–19:40.580
If you got some kind of
19:40.580–19:42.847
value out of this then please hit subscribe.
19:42.847–19:45.420
I would massively appreciate it and make sure that
19:45.420–19:48.698
you check out the pinned comment for all of the free resources.
19:48.698–19:50.400
You'll also find a full clod code
19:50.400–19:54.960
masterclass in the school which will take you from complete beginner to proficient
19:54.960–19:58.960
cloud code user so thanks for watching i'll see you in the next one
0:00.000–0:01.597
如果你正在觀看這段影片，那麼
0:01.597–0:03.140
我猜你應該是
0:03.140–0:06.280
某個在使用 Claude 訂閱時遇到方案限制的人。
0:06.880–0:10.305
好消息是，我接下來要向你展示 11 種方法，
0:10.305–0:13.600
讓你確保永遠不再遇到這些使用量限制。
0:13.860–0:17.669
而且有一個大多數人完全忽略的關鍵基本原則，
0:17.669–0:21.120
它將永遠改變你與 Claude Code 互動的方式。
0:21.760–0:27.260
所以你需要了解的第一件事是，Claude Code 實際上根本不記得任何事情。
0:27.260–0:31.844
因此，實際上運作的方式是，對於你發送到 CloudCode 的每一則新訊息，
0:31.844–0:36.631
它不僅接收最新的那則訊息，也會接收之前所有的內容
0:36.631–0:37.440
。
0:37.900–0:40.358
所以如果你在這個工作階段中與 CloudCode 進行了 50 次來回互動，
0:40.358–0:42.888
那麼
0:42.888–0:46.720
在第 51 則訊息時，它會再次接收所有的這些上下文。
0:47.220–0:48.813
你可能會想，嗯，
0:48.813–0:51.940
這會消耗數百萬個 token，而它確實會。
0:51.940–0:54.300
這就引出了快取的概念。
0:54.300–1:01.600
所以快取是 Claude Code 用來判斷哪些資訊已存在於工作階段中，以及哪些是新增資訊的方式。
1:02.020–1:07.680
這就是為什麼我們不會在短短 20 或 30 分鐘後就耗盡工作階段限制的原因。
1:08.260–1:12.011
所以在第一輪，也就是你在該工作階段中的第一次互動，
1:12.011–1:15.832
Claude 會收到系統提示詞、專案上下文
1:15.832–1:17.460
以及你的初始訊息。
1:17.780–1:18.868
然後在第二輪，
1:18.868–1:20.035
Claude 知道
1:20.035–1:24.000
來自第一輪的所有這些資訊已經存在。
1:24.300–1:28.811
所以在第二輪，你只需為 Claude 的回應和
1:28.811–1:30.240
你的下一則提示詞付費。
1:30.700–1:31.960
第三輪也是一樣的。
1:32.320–1:35.840
它接著知道來自第二輪的所有資訊已經存在。
1:36.200–1:39.360
所以你只需為最後的新內容付費。
1:39.360–1:42.414
而在此回覆之前發生的所有事情，
1:42.414–1:44.500
所有這些都被稱為前綴（prefix）。
1:44.840–1:47.887
稍後在討論技巧時，你會明白為什麼這如此重要，
1:47.887–1:49.120
但現在只需理解，如果快取資料中的某些內容或前綴中的某些內容發生變化
1:49.120–1:54.800
，
1:54.800–2:01.060
你就需要為所有這些內容再次付費，這意味著如果你的上下文視窗中有 50 萬或 60 萬
2:01.060–2:03.785
個 token，並且
2:03.785–2:06.320
前綴中的某些內容發生變化，那麼你
2:06.320–2:08.411
就需要為所有這些內容再次付費。所以
2:08.411–2:10.860
這一切的意思是，Claude 並不記得
2:10.860–2:12.498
回合之間，它不會記得你是誰。
2:12.498–2:16.520
每當發送新訊息時，它都會將所有內容重新發送給 Claude，
2:16.520–2:19.613
它會重新閱讀這些內容，知道哪些是原本就有的，
2:19.613–2:21.160
也知道哪些是新的，而你
2:21.160–2:24.498
只會為新增的部分支付全額費用。
2:24.498–2:27.060
現在，如果你是在使用 API，我們知道
2:27.060–2:31.468
快取輸入的成本僅為初始輸入成本的 0.1 倍。
2:31.468–2:34.700
Claude 並未在公開文件
2:34.700–2:38.773
中披露計劃訂閱的倍率是多少，
2:38.773–2:40.920
但我們確實知道，我們並不需要
2:40.920–2:42.694
為這些權杖支付全額費用。
2:42.694–2:44.320
同樣地，這將是
2:44.320–2:46.980
初始輸入成本的一定比例。
2:47.100–2:51.500
現在我們已經釐清了這個重要的區別，從這裡開始的一切都會變得更加合理，
2:51.500–2:56.920
所以讓我們進入第一個技巧，那就是將所有重要的上下文保留在你的
2:56.920–2:58.439
claude.md 檔案中。claude.
2:58.439–2:59.875
md 只是一個位於你專案內的
2:59.875–3:03.420
Markdown 檔案，它會在該專案的每個新工作階段開始時
3:03.420–3:09.460
自動注入你的上下文視窗中。所以
3:09.460–3:12.193
這之所以如此重要，是因為如果你沒有使用 claude.
3:12.193–3:14.660
md 檔案，那麼在每個新工作階段的開始，
3:14.660–3:17.967
你都需要解釋你是誰、你的技術堆疊是什麼、
3:17.967–3:20.340
你的檔案放在哪裡、你的規則為何，
3:20.340–3:23.955
然後你才需要給它任務或初始提示。
3:23.955–3:25.980
而如果你使用 claude.md，
3:25.980–3:28.636
那麼所有這些上下文已經儲存在那裡，所以
3:28.636–3:30.960
Claude 已經擁有所有這些資訊。
3:31.520–3:33.492
你不需要浪費權杖來來回回
3:33.492–3:36.160
每次都在解釋相同的事情。所以它不僅會
3:36.160–3:38.820
為你節省權杖，還會節省你的時間。
3:38.820–3:41.480
你可以在這裡看到我的專案中的 claude.md。
3:41.880–3:43.482
現在，如果你想建立一個，
3:43.482–3:46.620
你可以執行斜線指令 /init，你可以看到這裡，
3:47.060–3:48.383
初始化一個新的 claude.
3:48.383–3:49.706
md 檔案，包含程式碼庫
3:49.706–3:52.560
文件，或者你只需直接用
3:52.560–3:55.307
英文描述你想要它為你建立一個 claude.md，
3:55.307–3:57.400
它就會幫你完成。技巧
3:57.400–4:01.268
二是將你的上下文保持在 20% 到 25% 以下。
4:01.268–4:04.400
如果 Opus 4.8 和其他這些新的 Anthropic
4:04.400–4:07.486
所有模型都具備一百萬個 token 的上下文，那麼
4:07.486–4:10.340
為什麼我們要將內容控制在 25% 以下呢？
4:10.340–4:13.104
嗯，這裡有多種不同的原因，但
4:13.104–4:14.740
我現在只涵蓋其中幾點
4:14.740–4:20.280
此後的三個技巧都與這個相同的問題有關，因此這非常
4:20.280–4:24.295
非常重要。第一點是我在介紹中提到的，當
4:24.295–4:26.080
我們第一次向 Clod API 發送提示
4:26.080–4:31.880
時，在 Clod Code 中我們將為這些 token 支付 100% 的成本
4:31.880–4:35.260
所以我們是按每個 token 支付一個 token 的費用。
4:35.260–4:37.880
現在在那之後，一旦這些 token 被
4:37.880–4:40.029
緩存並成為前綴的一部分，
4:40.029–4:42.240
我們仍然在為這些 token 付費。而我們
4:42.240–4:47.160
不知道 Anthropic 訂閱或 Claude 方案的這些成本是多少，但我們知道
4:47.160–4:51.192
這些 token 不是免費的。所以即使是初始成本的 0.
4:51.192–4:54.820
05 倍，也就是每個 token 的 1/20，
4:55.240–4:58.557
如果你的上下文視窗中有 500,000 個 token，
4:58.557–5:01.400
那麼你每輪要支付 25,000 個 token，
5:01.400–5:05.864
並且這會隨著每次新輪次僅在緩存讀取中增加。
5:05.864–5:07.160
這是一個原因。
5:07.600–5:13.000
現在，另一個你想將上下文保持在 20% 以下的原因是，我們知道這些模型
5:13.000–5:17.378
隨著上下文填充得越多，性能開始下降。
5:17.378–5:18.720
這意味著如果
5:18.720–5:22.588
你有 800,000 個 token，即使用了上下文視窗的 80%，
5:22.588–5:24.600
那麼模型的表現
5:24.600–5:28.040
就不會像在該上下文視窗中只有 150,
5:28.040–5:30.840
000 個 token 時那樣達到最佳狀態。你可以
5:30.840–5:35.030
在我的狀態列中看到，我的上下文使用情況以條形圖和
5:35.030–5:36.280
百分比列出的形式
5:36.280–5:39.059
顯示，這意味著我可以隨時一目了然地看到。
5:39.059–5:41.700
現在我建議你也這樣做，我
5:41.700–5:44.207
會在螢幕上顯示一個提示，你可以複製並
5:44.207–5:46.400
交給 Cloud Code，讓它為
5:46.400–5:47.754
你實現相同的東西。
5:47.754–5:48.966
如果你想隨時關注
5:48.966–5:51.960
你的上下文，只需運行斜杠 context
5:51.960–5:55.095
它就會顯示這個視覺介面，讓你看到你所使用的內容。
5:55.095–5:56.520
你可以在這裡
5:56.520–5:59.251
右側看到它的分解，它也在下方進行了分解。
5:59.251–6:00.820
當你開始接近那個
6:00.820–6:05.886
20 到 25% 的限制時，我建議你要求 Claude 生成一個交接文件，
6:05.886–6:07.520
那只是一份 Markdown
6:07.520–6:11.095
檔案，裡面有該工作階段中所有已完成工作的摘要，
6:11.095–6:13.020
並提供你一個工作階段恢復
6:13.020–6:14.395
提示詞。這樣你就可以
6:14.395–6:15.841
將該工作階段恢復
6:15.841–6:18.880
提示詞帶回新工作階段中的 Claude，或者在
6:18.880–6:22.062
清除你的上下文後，讓它讀取交接檔案，
6:22.062–6:23.780
然後你就可以精確地從
6:23.780–6:26.913
你離開的地方繼續進行。如果你想要再進一步，
6:26.913–6:29.300
我將這功能建構為一個技能，我稱之為
6:29.300–6:32.971
交接，所以我只需要執行斜線交接，
6:32.971–6:34.660
它就會為我完成所有這些工作。
6:34.920–6:39.560
如果你想要取得該技能，你可以在下方的釘選
6:39.560–6:40.667
留言中找到影片中的所有資源。
6:40.667–6:41.709
第三個建議是
6:41.709–6:44.640
確保你的提示詞在上一個提示詞的一小時內
6:44.640–6:47.492
發出。這是我看到人們在
6:47.492–6:49.900
Claude Code 內部最常犯的
6:49.900–6:51.365
簡單且常見的錯誤，
6:51.365–6:52.829
它會消耗掉你的
6:52.829–6:55.480
五小時工作階段限制以及你 Claude 訂閱的
6:55.480–6:57.237
每週配額。所以
6:57.237–7:00.880
我們正在查看 Claude Code 的文件，你可以看到
7:00.880–7:04.851
它說 Claude Code 會自動請求一小時的 TTL（生存時間）。
7:04.851–7:06.440
所以我們在這裡再次談到
7:06.440–7:09.440
快取。因此，當你使用 Claude Code 時，
7:09.440–7:12.140
每次你發送新訊息時，都會
7:12.140–7:16.500
重置一小時的快取。現在如果你超過了一小時，
7:16.500–7:18.980
快取就會失效，
7:18.980–7:20.094
而你的下一個提示詞，
7:20.094–7:21.145
你將需要為那些
7:21.145–7:23.620
令牌再次付費。所以如果你有
7:23.620–7:25.893
500,000 個令牌坐在你的上下文視窗中，
7:25.893–7:28.280
如果你遵循第二步，你就不應該有這麼多，
7:28.280–7:31.543
那麼如果你離開一小時然後回來，
7:31.543–7:33.820
你發送下一個提示詞給 Claude，
7:34.160–7:37.772
你將為那 500,000 個令牌支付全額費用。
7:37.772–7:40.180
如果你再離開一小時，
7:40.180–7:42.530
然後回來並發送下一個提示詞，
7:42.530–7:45.640
你現在已經為那 500,000 個令牌支付了三次費用。
7:46.040–7:49.980
這就是為什麼你必須將對話保持在上一個提示的一小時內，這超級重要
7:49.980–7:54.380
並且回到之前的建議，將你的聯繫次數保持在20%以下
7:54.380–7:57.930
所以如果你超過了一小時，就不會每次都在燃燒500,
7:57.930–7:59.860
000個token。
8:02.020–8:03.106
在會話中途切換模型。
8:03.106–8:04.063
你有過多少次聽說你應該將模型從Sonnet
8:04.063–8:07.320
切換到Opus以處理更複雜的任務，然後
8:07.320–8:09.708
再切回Sonnet來處理較簡單的任務？現在我要告訴你，那是錯誤的建議。
8:09.708–8:13.140
我們停留在Cloud Code的文件中，你可以
8:13.140–8:15.243
在這裡看到會使快取無效的操作。
8:15.243–8:18.200
列表上的第一項是切換模型。所以如果我們
8:18.200–8:21.863
向下滾動，你可以看到它說每個模型都有自己的快取。
8:21.863–8:26.020
使用斜線
8:26.020–8:29.621
切換模型意味著下一個請求會讀取整個對話歷史，即使內容完全相同，也不會命中快取。所以
8:29.621–8:31.320
我看到人們一直在犯這個錯誤。他們不斷地在Opus、Sonnet、Sonnet、Opus之間
8:31.320–8:38.040
切換模型，
8:38.040–8:39.895
可能是Fable，而你每次切換時都要為這些
8:39.895–8:43.540
token再次付費。
8:43.540–8:47.020
我不是說你不應該
8:47.020–8:49.820
切換模型，因為
8:49.820–8:52.198
根據手頭任務使用正確的模型確實是好建議。我所說的是你不應該在會話中途切換模型。你應該
8:52.198–8:53.820
使用你更高階的模型制定計劃，
8:53.820–8:55.386
無論是Opus還是Fable，
8:55.386–8:58.460
然後一旦你有了結構化的計劃，
8:58.460–9:04.700
將你的模型切換到你要用來執行該任務的任何模型，並且
9:04.700–9:07.101
除非絕對必要，否則不要再切換。第五項幾乎相同，並且
9:07.101–9:08.980
那就是避免在會話中途改變努力等級。
9:09.440–9:11.753
就在切換模型和
9:11.753–9:13.840
使快取無效的操作下方，我們有改變
9:13.840–9:16.160
努力等級。所以
9:16.160–9:18.140
你可以在這裡看到
9:18.140–9:21.084
它說快取也是根據努力等級和
9:21.084–9:23.640
模型來鍵入的，所以使用斜線切換努力等級意味著下一個請求會再次讀取整個
9:23.640–9:26.657
歷史，不會命中快取。所以
9:26.657–9:29.900
使快取失效的操作，我們有改變
9:29.900–9:30.934
努力程度。所以
9:30.934–9:31.894
你可以在這裡看到
9:31.894–9:35.440
它說快取是根據努力程度以及
9:35.440–9:41.120
模型來索引的，因此使用斜杠和 effort 進行切換意味著下一個請求會重新讀取整個
9:41.120–9:43.660
歷史記錄，且沒有快取命中。所以
9:43.660–9:47.640
這與上一個建議完全相同，如果你能避免
9:47.640–9:49.816
在會話中途更改努力等級，那麼
9:49.816–9:53.080
你絕對應該這麼做。第六點是建立一個
9:53.080–9:54.273
完善的記憶系統。
9:54.273–9:55.277
我的頻道上有一段完整的影片
9:55.277–9:58.040
詳細說明如何設定這個系統
9:58.040–10:01.190
給你自己。我不會在這裡深入探討，但
10:01.190–10:03.680
我會將該影片連結顯示在螢幕上。不過我們
10:03.680–10:07.632
在這裡要討論的是建立一個完善的記憶系統，以便
10:07.632–10:08.880
Claude Code 擁有
10:08.880–10:10.551
我們專案的工作地圖，並且
10:10.551–10:13.960
它總是知道要去哪裡尋找我們所要求的
10:13.960–10:17.101
資訊。我們可以透過使用類似 Obsidian 的東西來做到這一點。
10:17.101–10:18.820
如果你不知道 Obsidian
10:18.820–10:21.753
是什麼，它只是一套讓我們能夠檢視和
10:21.753–10:24.400
編輯標記檔的軟體。你可以在這裡看到
10:24.400–10:29.280
我的 Obsidian 知識圖表範例，你會得到一個非常酷的圖表，我們可以像拉動
10:29.280–10:33.320
這些不同的節點一樣拉動這些不同的線條，你看到的這些線條都是
10:33.320–10:35.127
不同文件之間的連接，所以
10:35.127–10:38.340
它可以將不同部分的資訊組合起來，在
10:38.340–10:42.480
不同文件之間工作，以找到正確的資訊，為你提供高層次的概覽。
10:42.480–10:46.248
關於這如何運作，我們在電腦上的某個位置建立一個資料夾，
10:46.248–10:48.060
然後我們將 Obsidian 指向這個
10:48.060–10:49.677
資料夾，然後那個
10:49.677–10:51.294
就成為我們的保險庫。現在一旦
10:51.294–10:54.220
我們有了保險庫，我們可以將 Claude Code
10:54.220–10:55.305
指向這個保險庫，並且
10:55.305–10:56.390
在我們的保
10:56.390–11:00.420
險庫內部，我們將擁有很多不同的資料夾。現在這些
11:00.420–11:02.473
資料夾可以是任何你想要的樣子。
11:02.473–11:06.140
對我來說，我喜歡將這分為三個不同的層級。所以
11:06.140–11:08.627
第一層是我的工作資料夾。所以
11:08.627–11:12.880
對我來說，那是像機構、學校、內容這樣的事物，而在這些
11:12.880–11:15.167
資料夾內部，我會有子資料夾。所以
11:15.167–11:18.260
對於機構，那將是像客戶、它可能是
11:18.260–11:21.464
交付成果、外展、管線，對於學校，
11:21.464–11:24.980
它可能是框架、策略、課程。所以那是
11:24.980–11:25.923
我們的第一層，
11:25.923–11:26.799
所有工
11:26.799–11:30.100
作文件，你將在裡面每天進行操作
11:30.100–11:35.640
到日常基礎上。這裡的第二層源自 Andre Carpathie 的 Obsidian RAG 系統。我
11:35.640–11:37.077
不會在這裡詳細說明，所以
11:37.077–11:39.840
如果你對此感興趣，請務必查看螢幕上連結的那部影片。而第三層則是我們的 clod。
11:39.840–11:43.045
.md。所以正如我們在提示一
11:43.045–11:45.580
中提到的，這裡是我們儲存專案地圖、規則、
11:45.580–11:48.790
上下文的地方，以便 Claude 在會話開始時就知道所有
11:48.790–11:51.500
它需要了解關於我們的事情，而不需要我們每次都要解釋。那麼
11:51.500–11:56.200
這如何與我們這裡的記憶系統相關呢？首先，
11:56.200–11:59.454
我們給 Claude 一個專案資料夾的地圖，
11:59.454–12:02.000
每個資料夾裡包含什麼內容，以便
12:02.440–12:04.609
它知道東西在哪裡。然後我們需要給它一張地圖，
12:04.609–12:08.080
告訴它當我們要求儲存東西時，
12:08.080–12:09.982
應該把東西放在哪裡。所以你可以看到，這裡任何
12:09.982–12:13.000
屬於客戶的東西都放在 agency clients 和客戶名稱下。如果是想法、腳本或
12:13.000–12:15.814
鉤子，則放在 content 中。所以我們提供它所需的信息，以便
12:15.814–12:18.940
當它儲存這些東西時，
12:18.940–12:22.261
不會遺失。最後
12:22.261–12:23.120
我們還提供它知識庫規則。所以這是它
12:23.120–12:25.227
知道如何操作我們的知識庫系統的方式，
12:25.227–12:29.280
這 arguably 是最重要的部分。
12:29.280–12:31.953
你可以看到這裡有關於 wiki 連結的信息，以確保它始終在
12:31.953–12:34.180
不同文件之間建立
12:34.540–12:38.420
這些連結，這就是我們如何在
12:38.420–12:41.180
obsidian.md 內部獲得漂亮圖表的原因。第七點是建立自訂技能。
12:41.180–12:43.940
這一點毫無疑問是本影片中最強大的
12:43.940–12:47.550
提示，不僅從節省 token 的角度來看，
12:47.550–12:50.140
而且從最大化你
12:50.140–12:53.841
從 Claude Code 獲得的價值來看，特別是針對你個人。
12:53.841–12:55.840
所以建立自訂技能只是打包
12:55.840–12:59.265
一些關於特定主題的豐富信息，這些信息是針對你個人的，以便
12:59.265–13:02.300
它是可重複的，
13:02.300–13:07.420
並且 Claude 可以一遍又一遍地執行，
13:07.420–13:08.860
而不需要我們回去二十次，改變
13:08.860–13:10.393
這個，改變那個，因為
13:10.393–13:14.060
Claude 已經知道良好的輸出是什麼樣子。現在我強烈
13:14.060–13:15.718
建議你安裝 Anthropic 的技能創建者技能。
13:15.718–13:19.900
這是 Anthropic 的官方技能，
13:19.900–13:23.878
它告訴 Claude Code 如何構建自訂技能。
13:23.878–13:26.080
這是 Anthropic 的官方技能
13:26.080–13:29.038
它告訴 Claude Code 如何構建自定義技能。
13:29.038–13:31.240
所以首先，你要解釋
13:31.240–13:33.035
你想要做什麼。然後
13:33.035–13:35.320
Claude Code 會問你一堆問題，
13:35.720–13:37.914
它會測試它，執行評估，
13:37.914–13:39.160
它會確保它
13:39.160–13:40.434
在適當的時候觸發，並且
13:40.434–13:43.300
它會確保你對輸出結果滿意。所以接下來
13:43.300–13:44.497
當你想要使用它時，
13:44.497–13:45.694
你只需要
13:45.694–13:48.420
在未來執行斜線指令，然後這個任務
13:48.420–13:52.119
現在就可以重複使用了。這就是為什麼這如此
13:52.119–13:54.480
強大的原因，沒有技能時，你要求它
13:54.480–13:55.819
做某件事，然後你必須
13:55.819–13:59.020
回去說不，那不對，你可能需要一次又一次地回去，
13:59.020–14:00.579
一次又一次，一次又一次，
14:00.579–14:04.280
最終你會得到一個讓你滿意的產品，但到那個
14:04.280–14:06.408
時候，你已經消耗了大量 token，
14:06.408–14:09.100
如果你每週多次執行此操作，那麼我
14:09.100–14:14.840
可以保證你正在浪費大量計劃限制中的 token，用於可重複的任務。所以我
14:14.840–14:18.473
建議你複製螢幕上顯示的提示詞，並
14:18.473–14:19.900
讓 Claude 執行
14:19.900–14:21.869
你過去 10、15、20 個工作階段，並
14:21.869–14:25.660
挑出你多次要求它執行的不同事項，
14:25.660–14:28.397
並告訴它將這些事項建構為技能。
14:28.397–14:31.060
要安裝技能建立者技能，只需執行
14:31.060–14:33.063
斜線指令 plugins，然後
14:33.063–14:37.300
在左側的「探索」標籤頁中，你會看到它列在第四個。
14:37.440–14:40.244
你可以看到技能建立者建立新技能，
14:40.244–14:43.200
改進現有技能以及衡量技能。所以你所需要做的就是輸入這個並
14:43.200–14:45.796
安裝它。我建議你選擇
14:45.796–14:48.540
使用者範圍，這意味著你可以在所有不同的專案中存取。
14:48.540–14:52.540
第八點是使用 Caveman。
14:52.540–14:54.640
Caveman 是一個讓 Claude Code 減少冗長說明的外掛程式。
14:54.980–14:58.320
所以我們知道這些大型語言模型喜歡說話。
14:58.780–15:01.700
你問它一個簡單的問題，它會給你一個巨大的回覆，
15:02.040–15:04.880
這會花你五分鐘來閱讀。
15:05.200–15:06.820
而 Caveman 的設計旨在嘗試阻止這種情況發生。
15:07.440–15:10.600
而且它實際上做得非常好。
15:10.840–15:12.600
我已經斷斷續續地使用了它
15:12.980–15:14.028
大約三到四個月。
15:14.028–15:16.020
現在斷斷續續用了大概三到四個月。
15:16.280–15:17.700
它非常有效，
15:17.880–15:21.740
特別是如果你不希望從模型得到一大段文字回覆的話。
15:22.320–15:26.240
所以這裡寫著，為什麼要用很多 token，當少量就能解決時？
15:26.580–15:27.700
它的運作方式大致如此。
15:28.120–15:30.520
讓你的 AI 程式碼代理助手像穴居人一樣說話。
15:30.520–15:34.420
答案相同，輸出 token 減少高達 65%。
15:34.520–15:35.923
就我個人而言，
15:35.923–15:39.540
我可能還沒看到減少 65% 的輸出 token，
15:39.700–15:42.800
但它確實減少了 token 的輸出。
15:43.020–15:44.100
你可以在這裡看到一個範例。
15:44.220–15:48.200
我說，解釋提示緩存如何降低大型語言模型的費用。
15:48.200–15:51.900
這是使用 Caveman 時我得到的回覆。
15:52.360–15:54.260
現在如果我們將它與另一個終端進行比較，
15:54.740–15:56.140
我給了它完全相同的提示，
15:56.260–15:59.320
解釋提示緩存如何降低大型語言模型的費用，
15:59.320–16:00.000
以及……
16:00.000–16:01.174
降低大型語言模型的費用，以及
16:01.174–16:02.349
我從這裡得到的回覆
16:02.349–16:04.580
要長得多。你可以
16:04.580–16:07.018
看到這裡回覆大小的差異。所以
16:07.018–16:09.340
這絕對是你應該嘗試的
16:09.340–16:11.350
選項。第九個是使用 Ponytail。
16:11.350–16:14.700
這個工具在過去幾週被廣泛討論。你可以看到
16:14.700–16:17.345
它當天的排名第一，
16:17.345–16:20.060
當週也排名第一，它是為了這些 AI
16:20.060–16:23.672
程式碼代理助手喜歡寫不必要的大量程式碼而設計的，
16:23.672–16:26.200
Ponytail 基本上旨在讓
16:26.200–16:30.474
你的 AI 程式碼代理助手盡可能懶惰，同時不讓它懈怠。
16:30.474–16:31.560
所以如果我們查看
16:31.560–16:34.817
這裡的基準測試，你可以看到它的影響。
16:34.817–16:37.060
這是與無技能基準相比，
16:37.060–16:42.642
程式碼行數減少 54%，token 減少 22%，
16:42.642–16:48.080
費用減少 20%，時間減少 27%。如果
16:48.080–16:49.801
我們繼續往下滾動，
16:49.801–16:53.600
你可以看到它的運作層級結構。所以在寫任何
16:53.600–16:56.335
程式碼之前，代理助手會在第一個成立的階層停下。
16:56.335–16:58.340
它首先要做的是
16:58.340–17:01.605
檢查，這真的需要存在嗎？只有當答案是肯定的，
17:01.605–17:03.480
它才會接著看是否有
17:03.480–17:06.468
程式碼庫中是否已有能處理此任務的程式。
17:06.468–17:08.460
如果答案是否定的，它會接著檢視，
17:08.900–17:11.781
我能否用單一行程式碼來處理？如果
17:11.781–17:13.960
答案是否定的，它才會
17:13.960–17:16.850
編寫完成任務所需的最少量程式碼。
17:16.850–17:19.000
第 10 點是使用 CloudCode 的 Codex 外掛。
17:19.000–17:23.980
如果你已經訂閱了 OpenAI，這是一個非常棒的外掛。
17:23.980–17:29.700
它讓我們能在 CloudCode 生態系統中使用 Codex，
17:29.700–17:31.020
這意味著我們可以
17:31.020–17:35.415
將來自 Anthropic 模型的任務轉派給 GPT 模型。
17:35.415–17:38.000
現在我們知道這些 GPT 模型
17:38.000–17:40.867
比 Anthropic 模型便宜得多，
17:40.867–17:43.160
而且它們現在也非常具有競爭力。
17:43.700–17:45.341
GPT 5.6 剛剛發布，所以
17:45.341–17:48.380
如果你已經訂閱了 OpenAI 方案，
17:48.380–17:50.795
這絕對是你可以利用的好處。
17:50.795–17:53.080
安裝外掛後，只需輸入
17:53.080–17:54.187
斜線 codex，
17:54.187–17:57.120
你就會看到所有可用的斜線指令。
17:57.120–18:01.553
其中一個你絕對應該試試的是 Codex 對抗性審查，
18:01.553–18:02.440
它會
18:02.440–18:07.180
讓 Codex 代理程式深入檢視你的程式碼庫，看看能否找到任何資安
18:07.180–18:10.405
漏洞。這絕對是你應該試試的功能。
18:10.405–18:11.820
進去玩玩看，
18:11.820–18:14.279
這樣可以為你節省大量 token。
18:14.279–18:16.860
最後，第 11 個建議是將研究
18:16.860–18:20.673
密集型任務轉派給使用 Notebook LMPi 的 Google 伺服器。
18:20.673–18:24.220
Notebook LMPi 讓我們能使用 Google 伺服器
18:24.220–18:27.231
執行任何深度的網路研究。我們也可以
18:27.231–18:29.740
用它來生成資訊圖表和投影片
18:29.740–18:32.146
簡報，以及許多其他酷炫的功能，
18:32.146–18:34.120
而且我們可以在 Cloud Code 內的
18:34.120–18:35.611
命令列中完成這些操作。
18:35.611–18:36.937
其中一個主要的好
18:36.937–18:40.500
處是我們知道研究流程、
18:40.920–18:44.610
子代理程式、代理團隊，所有這些都非常消耗 token，
18:44.610–18:45.840
我們真的不想
18:45.840–18:49.428
在這些任務上消耗我們的 Cloud 訂閱額度，
18:49.428–18:51.680
這時我們可以將這些任務
18:51.680–18:55.781
基本上免費地轉派給 Google 伺服器，使用 Notebook LMPi，
18:55.781–18:57.560
其中一個巨大的好處
18:57.560–18:58.581
是因為
18:58.581–19:02.460
其底層使用的是 Gemini，它可以存取所有的 YouTube 內容
19:02.460–19:03.304
字幕檔
19:03.304–19:04.027
以及
19:04.027–19:07.100
所有那些通常僅憑 Claude Code 無法取得的豐富資訊
19:07.100–19:09.317
所以
19:09.317–19:13.340
我要求 Claude Code 使用 Notebook LMPi CLI 來執行
19:13.340–19:17.043
關於 GPT 5.6 SOL 與 Claude Fable 5 的研究
19:17.043–19:20.560
然後將其轉化為資訊圖表。所以它經歷了這裡的過程，而這是我們最終得到的結果。
19:20.560–19:23.516
經歷了這裡的過程，而這是我們最終得到的結果。
19:23.516–19:25.240
現在請注意，我不需要做任何
19:25.240–19:29.880
任何事。我沒有使用任何我的 clod tokens 或我的方案限制，並且
19:29.880–19:31.040
這就是我們最終得到的
19:31.040–19:33.252
結果。所以這是一個很棒的工具，而且
19:33.252–19:36.000
你絕對應該利用它來節省
19:36.000–19:39.481
大量的 tokens。這就是今天影片的全部內容。
19:39.481–19:40.580
如果你從中獲得了某種
19:40.580–19:42.847
價值，請點擊訂閱。
19:42.847–19:45.420
我會非常感激，並確保你
19:45.420–19:48.698
查看置頂評論中的所有免費資源。
19:48.698–19:50.400
你還會在學校中找到完整的 clod code
19:50.400–19:54.960
大師課程，它將帶你從完全初學者到熟練
19:54.960–19:58.960
cloud code 用戶，所以感謝觀看，我們下一部影片見
0:00.000–0:01.597
If you're watching this video, then
如果你正在觀看這段影片，那麼
0:01.597–0:03.140
I'm going to guess that you're some
我猜你應該是
0:03.140–0:06.280
body who's been running into plan limits on your Claude subscription.
某個在使用 Claude 訂閱時遇到方案限制的人。
0:06.880–0:10.305
Well, the good news for you is that I'm about to show you 11 ways so
好消息是，我接下來要向你展示 11 種方法，
0:10.305–0:13.600
you can make sure you never run into those usage limits again.
讓你確保永遠不再遇到這些使用量限制。
0:13.860–0:17.669
And there's a key fundamental principle that most people totally overlook,
而且有一個大多數人完全忽略的關鍵基本原則，
0:17.669–0:21.120
and it will change the way that you interact with Claude code forever.
它將永遠改變你與 Claude Code 互動的方式。
0:21.760–0:27.260
So the first thing which you need to understand is that Claude code does not actually remember anything at all.
所以你需要了解的第一件事是，Claude Code 實際上根本不記得任何事情。
0:27.260–0:31.844
So the way this actually works is for every new message you send away to CloudCode,
因此，實際上運作的方式是，對於你發送到 CloudCode 的每一則新訊息，
0:31.844–0:36.631
it's not only receiving that latest message, it's also receiving everything prior t
它不僅接收最新的那則訊息，也會接收之前所有的內容
0:36.631–0:37.440
o that as well.
。
0:37.900–0:40.358
So if you've had 50 interactions back and
所以如果你在這個工作階段中與 CloudCode 進行了 50 次來回互動，
0:40.358–0:42.888
forth with CloudCode in your session, the
那麼
0:42.888–0:46.720
n on the 51st message, it's receiving all of that context again.
在第 51 則訊息時，它會再次接收所有的這些上下文。
0:47.220–0:48.813
And you might be thinking, well,
你可能會想，嗯，
0:48.813–0:51.940
that would be chewing through millions of tokens, and it would.
這會消耗數百萬個 token，而它確實會。
0:51.940–0:54.300
And that brings us on to the idea of caching.
這就引出了快取的概念。
0:54.300–1:01.600
So caching is how Claude Code knows what information already existed in the session versus what is new.
所以快取是 Claude Code 用來判斷哪些資訊已存在於工作階段中，以及哪些是新增資訊的方式。
1:02.020–1:07.680
And this is what stops us from burning through our session limit after maybe only 20 or 30 minutes.
這就是為什麼我們不會在短短 20 或 30 分鐘後就耗盡工作階段限制的原因。
1:08.260–1:12.011
So on the first turn, so your first interaction in that session,
所以在第一輪，也就是你在該工作階段中的第一次互動，
1:12.011–1:15.832
Claude is going to receive the system prompt, the project contex
Claude 會收到系統提示詞、專案上下文
1:15.832–1:17.460
t and your initial message.
以及你的初始訊息。
1:17.780–1:18.868
Then on turn two,
然後在第二輪，
1:18.868–1:20.035
Claude knows that
Claude 知道
1:20.035–1:24.000
all of this information from the first turn already existed.
來自第一輪的所有這些資訊已經存在。
1:24.300–1:28.811
So all you're going to be charged for on turn two is Claude's response and
所以在第二輪，你只需為 Claude 的回應和
1:28.811–1:30.240
then your next prompt.
你的下一則提示詞付費。
1:30.700–1:31.960
And it's the same for turn three.
第三輪也是一樣的。
1:32.320–1:35.840
It then knows that all of the information from turn two was already there.
它接著知道來自第二輪的所有資訊已經存在。
1:36.200–1:39.360
So you're only being billed for the new stuff at the end.
所以你只需為最後的新內容付費。
1:39.360–1:42.414
And everything which happens before this reply,
而在此回覆之前發生的所有事情，
1:42.414–1:44.500
all of this is known as the prefix.
所有這些都被稱為前綴（prefix）。
1:44.840–1:47.887
You'll see in a second why this is so important when
稍後在討論技巧時，你會明白為什麼這如此重要，
1:47.887–1:49.120
we get into the tips.
但現在只需理解，如果快取資料中的某些內容或前綴中的某些內容發生變化
1:49.120–1:54.800
but for now just understand that if something in the cash data or something in the prefix changes
，
1:54.800–2:01.060
then you're going to be paying for all of this a second time which means if you have 500 or 600
你就需要為所有這些內容再次付費，這意味著如果你的上下文視窗中有 50 萬或 60 萬
2:01.060–2:03.785
thousand tokens sitting in your context window and
個 token，並且
2:03.785–2:06.320
something in the prefix has changed then you're
前綴中的某些內容發生變化，那麼你
2:06.320–2:08.411
going to be paying for all of that again. So
就需要為所有這些內容再次付費。所以
2:08.411–2:10.860
all of that is to say that Claude does not remember
這一切的意思是，Claude 並不記得
2:10.860–2:12.498
who you are between turns.
回合之間，它不會記得你是誰。
2:12.498–2:16.520
Every new message it's sending everything back to Claude again,
每當發送新訊息時，它都會將所有內容重新發送給 Claude，
2:16.520–2:19.613
it's going to read through it, it knows what was already there,
它會重新閱讀這些內容，知道哪些是原本就有的，
2:19.613–2:21.160
it knows what is new, and you're
也知道哪些是新的，而你
2:21.160–2:24.498
only going to be billed full price for the new stuff.
只會為新增的部分支付全額費用。
2:24.498–2:27.060
Now if you're using it on the API, we know
現在，如果你是在使用 API，我們知道
2:27.060–2:31.468
that cached inputs are 0.1x of the initial input cost.
快取輸入的成本僅為初始輸入成本的 0.1 倍。
2:31.468–2:34.700
CLODs do not disclose publicly in their
Claude 並未在公開文件
2:34.700–2:38.773
documentation what that multiple is for planned subscriptions,
中披露計劃訂閱的倍率是多少，
2:38.773–2:40.920
but we do know that we're not paying
但我們確實知道，我們並不需要
2:40.920–2:42.694
full price for those tokens.
為這些權杖支付全額費用。
2:42.694–2:44.320
Again it's going to be some
同樣地，這將是
2:44.320–2:46.980
form of fraction of the initial input cost.
初始輸入成本的一定比例。
2:47.100–2:51.500
Now that we've cleared up that important distinction everything from here on will make much more sense
現在我們已經釐清了這個重要的區別，從這裡開始的一切都會變得更加合理，
2:51.500–2:56.920
so let's move into the first tip which is keeping all of your important context inside of your
所以讓我們進入第一個技巧，那就是將所有重要的上下文保留在你的
2:56.920–2:58.439
clod.md file. A clod.
claude.md 檔案中。claude.
2:58.439–2:59.875
md is just a markdown
md 只是一個位於你專案內的
2:59.875–3:03.420
file which lives inside of your project and it gets
Markdown 檔案，它會在該專案的每個新工作階段開始時
3:03.420–3:09.460
auto-injected into your context window at every new session start in that project. So the reason
自動注入你的上下文視窗中。所以
3:09.460–3:12.193
this is so important is if you're not using a clod.
這之所以如此重要，是因為如果你沒有使用 claude.
3:12.193–3:14.660
md file then at the start of every new session
md 檔案，那麼在每個新工作階段的開始，
3:14.660–3:17.967
you need to explain who you are, what your tech stack is,
你都需要解釋你是誰、你的技術堆疊是什麼、
3:17.967–3:20.340
where your files go, what your rules are
你的檔案放在哪裡、你的規則為何，
3:20.340–3:23.955
and then you need to give it your task or your initial prompt.
然後你才需要給它任務或初始提示。
3:23.955–3:25.980
Whereas if you are using a clod.md
而如果你使用 claude.md，
3:25.980–3:28.636
all of this context is already stored in there so
那麼所有這些上下文已經儲存在那裡，所以
3:28.636–3:30.960
clod already has all of that information.
Claude 已經擁有所有這些資訊。
3:31.520–3:33.492
You're not wasting tokens going back and
你不需要浪費權杖來來回回
3:33.492–3:36.160
forth explaining the same thing every time. So it won't
每次都在解釋相同的事情。所以它不僅會
3:36.160–3:38.820
just save you tokens, it will also save you time.
為你節省權杖，還會節省你的時間。
3:38.820–3:41.480
You can see my clod.md inside of my project here.
你可以在這裡看到我的專案中的 claude.md。
3:41.880–3:43.482
Now if you want to create one,
現在，如果你想建立一個，
3:43.482–3:46.620
you can run slash command slash init and you can see here,
你可以執行斜線指令 /init，你可以看到這裡，
3:47.060–3:48.383
initialize a new clod.
初始化一個新的 claude.
3:48.383–3:49.706
md file with code-base
md 檔案，包含程式碼庫
3:49.706–3:52.560
d documentation or you can just describe in plain
文件，或者你只需直接用
3:52.560–3:55.307
English that you want it to create you a clod.md and
英文描述你想要它為你建立一個 claude.md，
3:55.307–3:57.400
it will go ahead and do that for you. Tip
它就會幫你完成。技巧
3:57.400–4:01.268
number two is keeping your context below 20 to 25%.
二是將你的上下文保持在 20% 到 25% 以下。
4:01.268–4:04.400
If Opus 4.8 and these other new anthropic
如果 Opus 4.8 和其他這些新的 Anthropic
4:04.400–4:07.486
models all have a million tokens of context then
所有模型都具備一百萬個 token 的上下文，那麼
4:07.486–4:10.340
why would we aim to keep things sub 25 percent
為什麼我們要將內容控制在 25% 以下呢？
4:10.340–4:13.104
well there's a multitude of different reasons for this but
嗯，這裡有多種不同的原因，但
4:13.104–4:14.740
i'll just cover a couple of them now
我現在只涵蓋其中幾點
4:14.740–4:20.280
and the next three tips after this one are all related back to this same issue so this is very
此後的三個技巧都與這個相同的問題有關，因此這非常
4:20.280–4:24.295
very important so the first one is as i mentioned in the intro when
非常重要。第一點是我在介紹中提到的，當
4:24.295–4:26.080
we send our prompts the first
我們第一次向 Clod API 發送提示
4:26.080–4:31.880
time around to the clod api in clod code we're being charged 100 of the cost for those tokens
時，在 Clod Code 中我們將為這些 token 支付 100% 的成本
4:31.880–4:35.260
So we're being billed one token for every token.
所以我們是按每個 token 支付一個 token 的費用。
4:35.260–4:37.880
Now after that, once those tokens are
現在在那之後，一旦這些 token 被
4:37.880–4:40.029
cashed and they become part of the prefix,
緩存並成為前綴的一部分，
4:40.029–4:42.240
we're still paying for those tokens. And we
我們仍然在為這些 token 付費。而我們
4:42.240–4:47.160
don't know what that cost is for Anthropic subscriptions or Claude plans, but we do know
不知道 Anthropic 訂閱或 Claude 方案的這些成本是多少，但我們知道
4:47.160–4:51.192
that those tokens are not free. So even if it's 0.
這些 token 不是免費的。所以即使是初始成本的 0.
4:51.192–4:54.820
05x of the initial cost, so 1 20th of a token,
05 倍，也就是每個 token 的 1/20，
4:55.240–4:58.557
if you have 500,000 tokens in your context window,
如果你的上下文視窗中有 500,000 個 token，
4:58.557–5:01.400
then you're paying 25,000 tokens per turn,
那麼你每輪要支付 25,000 個 token，
5:01.400–5:05.864
and that's going to be increasing with every new turn just in cached reads.
並且這會隨著每次新輪次僅在緩存讀取中增加。
5:05.864–5:07.160
So that's one reason.
這是一個原因。
5:07.600–5:13.000
Now another reason that you would want to keep your context sub 20% is we know that these models
現在，另一個你想將上下文保持在 20% 以下的原因是，我們知道這些模型
5:13.000–5:17.378
start degrading in performance the more you start to fill this context up.
隨著上下文填充得越多，性能開始下降。
5:17.378–5:18.720
So what that means is if
這意味著如果
5:18.720–5:22.588
you have 800,000 tokens, so 80% of your context window used,
你有 800,000 個 token，即使用了上下文視窗的 80%，
5:22.588–5:24.600
then the model is not going to be
那麼模型的表現
5:24.600–5:28.040
performing optimally as it would if we only had 150,
就不會像在該上下文視窗中只有 150,
5:28.040–5:30.840
000 tokens in that context window. You can
000 個 token 時那樣達到最佳狀態。你可以
5:30.840–5:35.030
see down here in my status line that I have my context usage as a bar and
在我的狀態列中看到，我的上下文使用情況以條形圖和
5:35.030–5:36.280
a percentage listed
百分比列出的形式
5:36.280–5:39.059
there which means I can always see it at a glance.
顯示，這意味著我可以隨時一目了然地看到。
5:39.059–5:41.700
Now I would recommend that you do the same and I
現在我建議你也這樣做，我
5:41.700–5:44.207
will put a prompt on screen which you can copy and
會在螢幕上顯示一個提示，你可以複製並
5:44.207–5:46.400
give to Cloud Code so that it can implement
交給 Cloud Code，讓它為
5:46.400–5:47.754
the same thing for you.
你實現相同的東西。
5:47.754–5:48.966
If you want to keep an
如果你想隨時關注
5:48.966–5:51.960
eye on your context just run forward slash context
你的上下文，只需運行斜杠 context
5:51.960–5:55.095
and it will bring up this visual where you can see what you've used.
它就會顯示這個視覺介面，讓你看到你所使用的內容。
5:55.095–5:56.520
You can see it broken down here
你可以在這裡
5:56.520–5:59.251
on the right hand side as well and it breaks it down below.
右側看到它的分解，它也在下方進行了分解。
5:59.251–6:00.820
When you start approaching that
當你開始接近那個
6:00.820–6:05.886
20 to 25% limit, I recommend that you ask Claude to generate a handoff file,
20 到 25% 的限制時，我建議你要求 Claude 生成一個交接文件，
6:05.886–6:07.520
which is just a markdown
那只是一份 Markdown
6:07.520–6:11.095
file with a summary of everything you've done in that session,
檔案，裡面有該工作階段中所有已完成工作的摘要，
6:11.095–6:13.020
and also give you a session resume
並提供你一個工作階段恢復
6:13.020–6:14.395
prompt. So you can then
提示詞。這樣你就可以
6:14.395–6:15.841
give that session resum
將該工作階段恢復
6:15.841–6:18.880
e prompt back to Claude in the new session, or after
提示詞帶回新工作階段中的 Claude，或者在
6:18.880–6:22.062
you've cleared your context, have it read the handoff file,
清除你的上下文後，讓它讀取交接檔案，
6:22.062–6:23.780
and you can pick up exactly where
然後你就可以精確地從
6:23.780–6:26.913
you left off. And if you want to go one step further,
你離開的地方繼續進行。如果你想要再進一步，
6:26.913–6:29.300
I built this into a skill, I just called
我將這功能建構為一個技能，我稱之為
6:29.300–6:32.971
it handoff so all I need to do is run forward slash handoff and
交接，所以我只需要執行斜線交接，
6:32.971–6:34.660
it will do all of that for me.
它就會為我完成所有這些工作。
6:34.920–6:39.560
If you want access to the skill you will find all of the resources from the video in the pinned
如果你想要取得該技能，你可以在下方的釘選
6:39.560–6:40.667
comment down below.
留言中找到影片中的所有資源。
6:40.667–6:41.709
Tip number three is
第三個建議是
6:41.709–6:44.640
making sure that you keep your prompts within one hour
確保你的提示詞在上一個提示詞的一小時內
6:44.640–6:47.492
of the previous prompt. This is one of the simplest and
發出。這是我看到人們在
6:47.492–6:49.900
most common mistakes that I see people making
Claude Code 內部最常犯的
6:49.900–6:51.365
inside of Claude Code and
簡單且常見的錯誤，
6:51.365–6:52.829
it will chew through your
它會消耗掉你的
6:52.829–6:55.480
five hour session limit and your weekly quota
五小時工作階段限制以及你 Claude 訂閱的
6:55.480–6:57.237
on your Claude subscription. So
每週配額。所以
6:57.237–7:00.880
we're looking at the Claude Code documentation here and you can see
我們正在查看 Claude Code 的文件，你可以看到
7:00.880–7:04.851
that it says Claude Code requests the one hour TTL automatically.
它說 Claude Code 會自動請求一小時的 TTL（生存時間）。
7:04.851–7:06.440
So we're talking about the
所以我們在這裡再次談到
7:06.440–7:09.440
caching again here. So when you use Claude Code,
快取。因此，當你使用 Claude Code 時，
7:09.440–7:12.140
every time you send a new message away, that
每次你發送新訊息時，都會
7:12.140–7:16.500
resets the one hour cache. Now if you go out with that one hour,
重置一小時的快取。現在如果你超過了一小時，
7:16.500–7:18.980
then the cache is then invalidated
快取就會失效，
7:18.980–7:20.094
and your next prompt,
而你的下一個提示詞，
7:20.094–7:21.145
you're going to be pa
你將需要為那些
7:21.145–7:23.620
ying for all of those tokens again. So if you have
令牌再次付費。所以如果你有
7:23.620–7:25.893
500,000 tokens sitting in your context window,
500,000 個令牌坐在你的上下文視窗中，
7:25.893–7:28.280
which you shouldn't have if you're following step
如果你遵循第二步，你就不應該有這麼多，
7:28.280–7:31.543
two, then if you go away for an hour and you come back,
那麼如果你離開一小時然後回來，
7:31.543–7:33.820
you send your next prompt to Claude,
你發送下一個提示詞給 Claude，
7:34.160–7:37.772
you're paying full price for those 500,000 tokens again.
你將為那 500,000 個令牌支付全額費用。
7:37.772–7:40.180
If you then go away for another hour and
如果你再離開一小時，
7:40.180–7:42.530
come back and then send your next prompt,
然後回來並發送下一個提示詞，
7:42.530–7:45.640
you've now paid for those 500,000 tokens three times.
你現在已經為那 500,000 個令牌支付了三次費用。
7:46.040–7:49.980
So this is why it's super important that you keep it within an hour of your previous prompt
這就是為什麼你必須將對話保持在上一個提示的一小時內，這超級重要
7:49.980–7:54.380
and going back to the previous tip that you keep your contacts below 20%
並且回到之前的建議，將你的聯繫次數保持在20%以下
7:54.380–7:57.930
so if you do go out with that hour you're not burning 500,
所以如果你超過了一小時，就不會每次都在燃燒500,
7:57.930–7:59.860
000 tokens every single time.
000個token。
8:02.020–8:03.106
model mid-session.
在會話中途切換模型。
8:03.106–8:04.063
How many times hav
你有過多少次聽說你應該將模型從Sonnet
8:04.063–8:07.320
e you heard that you should be changing your model from Sonnet
切換到Opus以處理更複雜的任務，然後
8:07.320–8:09.708
to Opus for more complex tasks and then
再切回Sonnet來處理較簡單的任務？現在我要告訴你，那是錯誤的建議。
8:09.708–8:13.140
back to Sonnet again for the easier stuff? And I'm going
我們停留在Cloud Code的文件中，你可以
8:13.140–8:15.243
to tell you now that that is bad advice.
在這裡看到會使快取無效的操作。
8:15.243–8:18.200
We're staying inside of the Cloud Code docs and you can
列表上的第一項是切換模型。所以如果我們
8:18.200–8:21.863
see here actions that invalidate the cache.
向下滾動，你可以看到它說每個模型都有自己的快取。
8:21.863–8:26.020
Number one on the list is switching models. So if we
使用斜線
8:26.020–8:29.621
scroll down you can see that it says each model has its own cache.
切換模型意味著下一個請求會讀取整個對話歷史，即使內容完全相同，也不會命中快取。所以
8:29.621–8:31.320
Switching with forward slash
我看到人們一直在犯這個錯誤。他們不斷地在Opus、Sonnet、Sonnet、Opus之間
8:31.320–8:38.040
model means the next request reads the entire conversation history with no cache hits even
切換模型，
8:38.040–8:39.895
know the content is identical. So
可能是Fable，而你每次切換時都要為這些
8:39.895–8:43.540
I see people making this mistake all the time. They're constantly
token再次付費。
8:43.540–8:47.020
switching models between Opus, Sonnet, Sonnet, Opus,
我不是說你不應該
8:47.020–8:49.820
maybe Fable, and you're paying for all those
切換模型，因為
8:49.820–8:52.198
tokens again every single time you make that switch.
根據手頭任務使用正確的模型確實是好建議。我所說的是你不應該在會話中途切換模型。你應該
8:52.198–8:53.820
I'm not saying that you should not be
使用你更高階的模型制定計劃，
8:53.820–8:55.386
switching models at all because
無論是Opus還是Fable，
8:55.386–8:58.460
that is genuinely good advice to use the right model based on the
然後一旦你有了結構化的計劃，
8:58.460–9:04.700
task at hand. What I am saying is that you should not be switching model mid-session. You should be
將你的模型切換到你要用來執行該任務的任何模型，並且
9:04.700–9:07.101
putting together a plan using your higher power model,
除非絕對必要，否則不要再切換。第五項幾乎相同，並且
9:07.101–9:08.980
whether that's Opus or whether it's Fable,
那就是避免在會話中途改變努力等級。
9:09.440–9:11.753
and then once you have a structured plan in place,
就在切換模型和
9:11.753–9:13.840
switch your model down to whatever it is that
使快取無效的操作下方，我們有改變
9:13.840–9:16.160
you're going to be using to execute that task, and
努力等級。所以
9:16.160–9:18.140
do not switch again unless you absolutely
你可以在這裡看到
9:18.140–9:21.084
have to. Number five is almost identical, and
它說快取也是根據努力等級和
9:21.084–9:23.640
that is avoiding changing effort level
模型來鍵入的，所以使用斜線切換努力等級意味著下一個請求會再次讀取整個
9:23.640–9:26.657
mid-session. Right below switching models and
歷史，不會命中快取。所以
9:26.657–9:29.900
actions that invalidate the cash, we have changing
使快取失效的操作，我們有改變
9:29.900–9:30.934
effort level. So
努力程度。所以
9:30.934–9:31.894
you can see here
你可以在這裡看到
9:31.894–9:35.440
that it says that the cache is keyed by effort level as well
它說快取是根據努力程度以及
9:35.440–9:41.120
as model so switching with forward slash effort means that the next request reads the entire
模型來索引的，因此使用斜杠和 effort 進行切換意味著下一個請求會重新讀取整個
9:41.120–9:43.660
history again with no cache hits. So
歷史記錄，且沒有快取命中。所以
9:43.660–9:47.640
this is the exact same as the previous tip if you can avoid
這與上一個建議完全相同，如果你能避免
9:47.640–9:49.816
changing effort level mid-session then
在會話中途更改努力等級，那麼
9:49.816–9:53.080
you should absolutely be doing that. Number six is building a
你絕對應該這麼做。第六點是建立一個
9:53.080–9:54.273
proper memory system.
完善的記憶系統。
9:54.273–9:55.277
I have a full video o
我的頻道上有一段完整的影片
9:55.277–9:58.040
n my channel breaking down how you can set this system
詳細說明如何設定這個系統
9:58.040–10:01.190
up for yourself. I'm not going to go too deep here but
給你自己。我不會在這裡深入探討，但
10:01.190–10:03.680
I will link that video on screen. But what
我會將該影片連結顯示在螢幕上。不過我們
10:03.680–10:07.632
we're talking about doing here is building a proper memory system so
在這裡要討論的是建立一個完善的記憶系統，以便
10:07.632–10:08.880
that Claude Code has a
Claude Code 擁有
10:08.880–10:10.551
working map of our project and
我們專案的工作地圖，並且
10:10.551–10:13.960
it always knows where to go to find the information that we're
它總是知道要去哪裡尋找我們所要求的
10:13.960–10:17.101
asking for. And we can do that by using something like Obsidian.
資訊。我們可以透過使用類似 Obsidian 的東西來做到這一點。
10:17.101–10:18.820
And if you don't know what Obsidian
如果你不知道 Obsidian
10:18.820–10:21.753
is, it's just a software that allows us to view and
是什麼，它只是一套讓我們能夠檢視和
10:21.753–10:24.400
edit marked-in files. And you can see here an
編輯標記檔的軟體。你可以在這裡看到
10:24.400–10:29.280
example of my obsidian knowledge graph you get this really cool graph we can kind of like pull
我的 Obsidian 知識圖表範例，你會得到一個非常酷的圖表，我們可以像拉動
10:29.280–10:33.320
around all these different nodes all these different lines that you see are connections
這些不同的節點一樣拉動這些不同的線條，你看到的這些線條都是
10:33.320–10:35.127
between different documents so
不同文件之間的連接，所以
10:35.127–10:38.340
it can piece together different bits of information work
它可以將不同部分的資訊組合起來，在
10:38.340–10:42.480
between different documents to find the right information to give you a high level overview of
不同文件之間工作，以找到正確的資訊，為你提供高層次的概覽。
10:42.480–10:46.248
how this works we create a folder somewhere on our computer and
關於這如何運作，我們在電腦上的某個位置建立一個資料夾，
10:46.248–10:48.060
we then point obsidian at this
然後我們將 Obsidian 指向這個
10:48.060–10:49.677
folder and then that then
資料夾，然後那個
10:49.677–10:51.294
becomes our vault now onc
就成為我們的保險庫。現在一旦
10:51.294–10:54.220
e we have our vault we can then point cloud code
我們有了保險庫，我們可以將 Claude Code
10:54.220–10:55.305
at this vault and
指向這個保險庫，並且
10:55.305–10:56.390
inside of our vau
在我們的保
10:56.390–11:00.420
lt we will have a bunch of different folders in there. Now these
險庫內部，我們將擁有很多不同的資料夾。現在這些
11:00.420–11:02.473
folders can be whatever you like.
資料夾可以是任何你想要的樣子。
11:02.473–11:06.140
For me I like to split this into three separate layers. So the
對我來說，我喜歡將這分為三個不同的層級。所以
11:06.140–11:08.627
first layer is my working folders. So
第一層是我的工作資料夾。所以
11:08.627–11:12.880
for me that's things like agency, school, content and inside of
對我來說，那是像機構、學校、內容這樣的事物，而在這些
11:12.880–11:15.167
these folders I will have subfolders. So
資料夾內部，我會有子資料夾。所以
11:15.167–11:18.260
for agency that will be things like clients, it could be
對於機構，那將是像客戶、它可能是
11:18.260–11:21.464
deliverables, outreach, pipeline, for school,
交付成果、外展、管線，對於學校，
11:21.464–11:24.980
it might be frameworks, strategy, courses. So that is
它可能是框架、策略、課程。所以那是
11:24.980–11:25.923
our first layer,
我們的第一層，
11:25.923–11:26.799
all of the worki
所有工
11:26.799–11:30.100
ng files that you'll be in there messing around with on a day
作文件，你將在裡面每天進行操作
11:30.100–11:35.640
to day basis. The second layer here originates from Andre Carpathie's Obsidian RAG system. I'm
到日常基礎上。這裡的第二層源自 Andre Carpathie 的 Obsidian RAG 系統。我
11:35.640–11:37.077
not going to go into that here so
不會在這裡詳細說明，所以
11:37.077–11:39.840
make sure you check out that video linked on screen if you're
如果你對此感興趣，請務必查看螢幕上連結的那部影片。而第三層則是我們的 clod。
11:39.840–11:43.045
interested in that. And the third layer is our clod.
.md。所以正如我們在提示一
11:43.045–11:45.580
md. So in there as we covered in tip number
中提到的，這裡是我們儲存專案地圖、規則、
11:45.580–11:48.790
one. This is where we store our project map, our rules,
上下文的地方，以便 Claude 在會話開始時就知道所有
11:48.790–11:51.500
our context, so that Claude knows everything
它需要了解關於我們的事情，而不需要我們每次都要解釋。那麼
11:51.500–11:56.200
which it needs to know about us at session start without us having to explain every time. So how
這如何與我們這裡的記憶系統相關呢？首先，
11:56.200–11:59.454
that relates to our memory system here is, first of all,
我們給 Claude 一個專案資料夾的地圖，
11:59.454–12:02.000
we give Claude a map of the project folders,
每個資料夾裡包含什麼內容，以便
12:02.440–12:04.609
what's contained in each folder, so
它知道東西在哪裡。然後我們需要給它一張地圖，
12:04.609–12:08.080
that it knows where things are. Then we need to give it a map
告訴它當我們要求儲存東西時，
12:08.080–12:09.982
of where it needs to put things when
應該把東西放在哪裡。所以你可以看到，這裡任何
12:09.982–12:13.000
we ask for them to be saved. So you can see here, anything
屬於客戶的東西都放在 agency clients 和客戶名稱下。如果是想法、腳本或
12:13.000–12:15.814
for a client goes in agency clients and then
鉤子，則放在 content 中。所以我們提供它所需的信息，以便
12:15.814–12:18.940
the client's name. If it's an idea or a script or a
當它儲存這些東西時，
12:18.940–12:22.261
hook it goes in content. So we give it the information which it needs so
不會遺失。最後
12:22.261–12:23.120
that things don't
我們還提供它知識庫規則。所以這是它
12:23.120–12:25.227
get lost when it saves them. And
知道如何操作我們的知識庫系統的方式，
12:25.227–12:29.280
finally we give it the knowledge base rules. So this is how it
這 arguably 是最重要的部分。
12:29.280–12:31.953
knows how to operate our knowledge base system and
你可以看到這裡有關於 wiki 連結的信息，以確保它始終在
12:31.953–12:34.180
this is arguably the most important part.
不同文件之間建立
12:34.540–12:38.420
You can see here we have information on wiki links to make sure that it's always creating
這些連結，這就是我們如何在
12:38.420–12:41.180
those links between different documents and
obsidian.md 內部獲得漂亮圖表的原因。第七點是建立自訂技能。
12:41.180–12:43.940
that is how we get our nice graph here inside of
這一點毫無疑問是本影片中最強大的
12:43.940–12:47.550
obsidian.md. Number seven is creating custom skills.
提示，不僅從節省 token 的角度來看，
12:47.550–12:50.140
This one is hands down the most powerful
而且從最大化你
12:50.140–12:53.841
tip in this video, not only from a token saving perspective,
從 Claude Code 獲得的價值來看，特別是針對你個人。
12:53.841–12:55.840
but also in maximizing the value
所以建立自訂技能只是打包
12:55.840–12:59.265
which you get out of cloud code for you specifically.
一些關於特定主題的豐富信息，這些信息是針對你個人的，以便
12:59.265–13:02.300
So creating a custom skill is just packaging up
它是可重複的，
13:02.300–13:07.420
some amount of rich information on a specific topic that's specific to you so
並且 Claude 可以一遍又一遍地執行，
13:07.420–13:08.860
that it's repeatable
而不需要我們回去二十次，改變
13:08.860–13:10.393
and Claude can do it over and
這個，改變那個，因為
13:10.393–13:14.060
over again without us having to go back 20 different times, change
Claude 已經知道良好的輸出是什麼樣子。現在我強烈
13:14.060–13:15.718
this, change that, because
建議你安裝 Anthropic 的技能創建者技能。
13:15.718–13:19.900
Claude already knows what a good output looks like. Now I would highly
這是 Anthropic 的官方技能，
13:19.900–13:23.878
recommend that you install the skill creator skill by Anthropics.
它告訴 Claude Code 如何構建自訂技能。
13:23.878–13:26.080
This is an official Anthropics skill
這是 Anthropic 的官方技能
13:26.080–13:29.038
and it tells Claude Code how to build custom skills.
它告訴 Claude Code 如何構建自定義技能。
13:29.038–13:31.240
So first of all you're going to explain
所以首先，你要解釋
13:31.240–13:33.035
what it is that you're trying to do. Then
你想要做什麼。然後
13:33.035–13:35.320
Cloud Code is going to ask you a bunch of questions,
Claude Code 會問你一堆問題，
13:35.720–13:37.914
it's going to test it, it's going to run evaluations,
它會測試它，執行評估，
13:37.914–13:39.160
it's going to make sure that it
它會確保它
13:39.160–13:40.434
triggers when it should, and
在適當的時候觸發，並且
13:40.434–13:43.300
it's going to make sure that you're happy with the output. So then
它會確保你對輸出結果滿意。所以接下來
13:43.300–13:44.497
all you need to do when
當你想要使用它時，
13:44.497–13:45.694
you want to use it goin
你只需要
13:45.694–13:48.420
g forward is run a slash command and then that task
在未來執行斜線指令，然後這個任務
13:48.420–13:52.119
is now repeatable on the mat. So the reason that this is so
現在就可以重複使用了。這就是為什麼這如此
13:52.119–13:54.480
powerful is without a skill, you ask
強大的原因，沒有技能時，你要求它
13:54.480–13:55.819
it to do something, you then
做某件事，然後你必須
13:55.819–13:59.020
have to go back saying no that's not right, you might need to go back
回去說不，那不對，你可能需要一次又一次地回去，
13:59.020–14:00.579
again and again and again and
一次又一次，一次又一次，
14:00.579–14:04.280
eventually you will have a product that you're happy with but by that
最終你會得到一個讓你滿意的產品，但到那個
14:04.280–14:06.408
point you've burned a bunch of tokens and
時候，你已經消耗了大量 token，
14:06.408–14:09.100
if you're doing this multiple times a week then I can
如果你每週多次執行此操作，那麼我
14:09.100–14:14.840
guarantee you're wasting a huge amount of your plan limits on tasks that are repeatable. So what I
可以保證你正在浪費大量計劃限制中的 token，用於可重複的任務。所以我
14:14.840–14:18.473
would suggest is that you copy the prompt which you see on screen and
建議你複製螢幕上顯示的提示詞，並
14:18.473–14:19.900
you have Claude go through
讓 Claude 執行
14:19.900–14:21.869
your last 10, 15, 20 sessions and
你過去 10、15、20 個工作階段，並
14:21.869–14:25.660
pick out things which you've asked it to do multiple different
挑出你多次要求它執行的不同事項，
14:25.660–14:28.397
times and tell it to build those into skills.
並告訴它將這些事項建構為技能。
14:28.397–14:31.060
To install the skill creator skill just run
要安裝技能建立者技能，只需執行
14:31.060–14:33.063
forward slash plugins and then
斜線指令 plugins，然後
14:33.063–14:37.300
on the discover tab here you will see it listed fourth from the top.
在左側的「探索」標籤頁中，你會看到它列在第四個。
14:37.440–14:40.244
You can see skill creator create new skills,
你可以看到技能建立者建立新技能，
14:40.244–14:43.200
improve existing skills and measure skills as
改進現有技能以及衡量技能。所以你所需要做的就是輸入這個並
14:43.200–14:45.796
well. So all you need to do is enter this and
安裝它。我建議你選擇
14:45.796–14:48.540
install it. I would recommend that you go for
使用者範圍，這意味著你可以在所有不同的專案中存取。
14:48.540–14:52.540
user scope which means you'll have access to across all of your different projects.
第八點是使用 Caveman。
14:52.540–14:54.640
Number eight is using Caveman.
Caveman 是一個讓 Claude Code 減少冗長說明的外掛程式。
14:54.980–14:58.320
Caveman is a plugin that makes Cloud Code less verbose.
所以我們知道這些大型語言模型喜歡說話。
14:58.780–15:01.700
So we know that these large language models love to talk.
你問它一個簡單的問題，它會給你一個巨大的回覆，
15:02.040–15:04.880
You ask it a simple question and it gives you a huge reply,
這會花你五分鐘來閱讀。
15:05.200–15:06.820
which will take you five minutes to read.
而 Caveman 的設計旨在嘗試阻止這種情況發生。
15:07.440–15:10.600
And Caveman was designed to try and stop that from happening.
而且它實際上做得非常好。
15:10.840–15:12.600
And it actually does a very, very good job.
我已經斷斷續續地使用了它
15:12.980–15:14.028
I've been using it on and
大約三到四個月。
15:14.028–15:16.020
off now for around about three or four months.
現在斷斷續續用了大概三到四個月。
15:16.280–15:17.700
And it is very effective,
它非常有效，
15:17.880–15:21.740
especially if you're not wanting a big walls of text back from the model.
特別是如果你不希望從模型得到一大段文字回覆的話。
15:22.320–15:26.240
So it says here, why use many token when few do trick?
所以這裡寫著，為什麼要用很多 token，當少量就能解決時？
15:26.580–15:27.700
And that is kind of how it works.
它的運作方式大致如此。
15:28.120–15:30.520
Make your AI coding agent talk like a caveman.
讓你的 AI 程式碼代理助手像穴居人一樣說話。
15:30.520–15:34.420
Same answers, up to 65% fewer output tokens.
答案相同，輸出 token 減少高達 65%。
15:34.520–15:35.923
Now for me personally,
就我個人而言，
15:35.923–15:39.540
I probably haven't seen as much as 65% fewer output tokens,
我可能還沒看到減少 65% 的輸出 token，
15:39.700–15:42.800
but it does definitely reduce the token output.
但它確實減少了 token 的輸出。
15:43.020–15:44.100
You can see an example here.
你可以在這裡看到一個範例。
15:44.220–15:48.200
I said, explain how prompt caching reduces cost in LLMs.
我說，解釋提示緩存如何降低大型語言模型的費用。
15:48.200–15:51.900
and this was the response that I got using Caveman.
這是使用 Caveman 時我得到的回覆。
15:52.360–15:54.260
Now if we compare that to this other terminal,
現在如果我們將它與另一個終端進行比較，
15:54.740–15:56.140
I gave it the exact same prompt,
我給了它完全相同的提示，
15:56.260–15:59.320
explain how prompt caching reduces cost in LLMs,
解釋提示緩存如何降低大型語言模型的費用，
15:59.320–16:00.000
and...
以及……
16:00.000–16:01.174
reduces cost in LLMs and
降低大型語言模型的費用，以及
16:01.174–16:02.349
the response which I got
我從這裡得到的回覆
16:02.349–16:04.580
from this one was much much longer. You can see
要長得多。你可以
16:04.580–16:07.018
the difference here in the size of the response. So
看到這裡回覆大小的差異。所以
16:07.018–16:09.340
this is definitely one that you should be trying
這絕對是你應該嘗試的
16:09.340–16:11.350
out. Number nine is using ponytail.
選項。第九個是使用 Ponytail。
16:11.350–16:14.700
This one has been blown up over the past few weeks. You can see
這個工具在過去幾週被廣泛討論。你可以看到
16:14.700–16:17.345
that it's number one repository of the day and
它當天的排名第一，
16:17.345–16:20.060
of the week and it was designed because these AI
當週也排名第一，它是為了這些 AI
16:20.060–16:23.672
coding agents love writing unnecessary amounts of code and
程式碼代理助手喜歡寫不必要的大量程式碼而設計的，
16:23.672–16:26.200
ponytail is essentially designed to make
Ponytail 基本上旨在讓
16:26.200–16:30.474
your AI coding agent as lazy as possible without making it complacent.
你的 AI 程式碼代理助手盡可能懶惰，同時不讓它懈怠。
16:30.474–16:31.560
So if we look at the
所以如果我們查看
16:31.560–16:34.817
benchmarks here you can see the impact which this has.
這裡的基準測試，你可以看到它的影響。
16:34.817–16:37.060
So this is versus a no skill baseline,
這是與無技能基準相比，
16:37.060–16:42.642
minus 54% on lines of code, minus 22% on tokens,
程式碼行數減少 54%，token 減少 22%，
16:42.642–16:48.080
minus 20% on cost, and minus 27% on time. And if
費用減少 20%，時間減少 27%。如果
16:48.080–16:49.801
we keep scrolling down here,
我們繼續往下滾動，
16:49.801–16:53.600
you can see this hierarchy of how it works. So before writing any
你可以看到它的運作層級結構。所以在寫任何
16:53.600–16:56.335
code, the agent will stop at the first rung that holds.
程式碼之前，代理助手會在第一個成立的階層停下。
16:56.335–16:58.340
The first thing which it's going to do is
它首先要做的是
16:58.340–17:01.605
check, does this even need to exist? Only if that is a yes,
檢查，這真的需要存在嗎？只有當答案是肯定的，
17:01.605–17:03.480
is it then going to see if there's
它才會接著看是否有
17:03.480–17:06.468
something already in the code base that can handle it.
程式碼庫中是否已有能處理此任務的程式。
17:06.468–17:08.460
And if that is a no, it's going to see,
如果答案是否定的，它會接著檢視，
17:08.900–17:11.781
can I handle this with one single line of code? And
我能否用單一行程式碼來處理？如果
17:11.781–17:13.960
only if that is a no, is it then going to
答案是否定的，它才會
17:13.960–17:16.850
write the minimum amount of code to get the job done.
編寫完成任務所需的最少量程式碼。
17:16.850–17:19.000
Number 10 is using the Codex plugin for
第 10 點是使用 CloudCode 的 Codex 外掛。
17:19.000–17:23.980
CloudCode. This is an amazing plugin if you're someone who's already paying for an OpenAI
如果你已經訂閱了 OpenAI，這是一個非常棒的外掛。
17:23.980–17:29.700
subscription. So it allows us to use Codex inside of the CloudCode ecosystem,
它讓我們能在 CloudCode 生態系統中使用 Codex，
17:29.700–17:31.020
which means we can
這意味著我們可以
17:31.020–17:35.415
offload tasks from the Anthropic models onto the GPT models.
將來自 Anthropic 模型的任務轉派給 GPT 模型。
17:35.415–17:38.000
Now we know that these GPT models are
現在我們知道這些 GPT 模型
17:38.000–17:40.867
substantially cheaper than the Anthropic models and
比 Anthropic 模型便宜得多，
17:40.867–17:43.160
they're now very very competitive as well.
而且它們現在也非常具有競爭力。
17:43.700–17:45.341
GPT 5.6 has just been released so
GPT 5.6 剛剛發布，所以
17:45.341–17:48.380
that is definitely something that you can take advantage of
如果你已經訂閱了 OpenAI 方案，
17:48.380–17:50.795
if you're already paying for an OpenAI plan.
這絕對是你可以利用的好處。
17:50.795–17:53.080
Once you have the plugin installed just do
安裝外掛後，只需輸入
17:53.080–17:54.187
forward slash codex and
斜線 codex，
17:54.187–17:57.120
you will see all the different slash commands which you now have
你就會看到所有可用的斜線指令。
17:57.120–18:01.553
available. One which you should definitely check out is Codex Adversarial Review,
其中一個你絕對應該試試的是 Codex 對抗性審查，
18:01.553–18:02.440
which is going to
它會
18:02.440–18:07.180
have the Codex agent do a deep dive into your code base to see if it can find any security
讓 Codex 代理程式深入檢視你的程式碼庫，看看能否找到任何資安
18:07.180–18:10.405
vulnerabilities. This is definitely one that you should check out.
漏洞。這絕對是你應該試試的功能。
18:10.405–18:11.820
Go in there, have a play around
進去玩玩看，
18:11.820–18:14.279
with it, and this will save you a bunch of tokens.
這樣可以為你節省大量 token。
18:14.279–18:16.860
And finally, tip number 11 is offloading research
最後，第 11 個建議是將研究
18:16.860–18:20.673
heavy tasks to Google servers using Notebook LMPi.
密集型任務轉派給使用 Notebook LMPi 的 Google 伺服器。
18:20.673–18:24.220
Notebook LMPi allows us to use Google servers to
Notebook LMPi 讓我們能使用 Google 伺服器
18:24.220–18:27.231
carry out any kind of deep web research. We can also
執行任何深度的網路研究。我們也可以
18:27.231–18:29.740
use it to generate infographics and slide
用它來生成資訊圖表和投影片
18:29.740–18:32.146
decks and a bunch of other cool stuff as well and
簡報，以及許多其他酷炫的功能，
18:32.146–18:34.120
we can do this all from the command line
而且我們可以在 Cloud Code 內的
18:34.120–18:35.611
inside of Cloud Code.
命令列中完成這些操作。
18:35.611–18:36.937
Now one of the main b
其中一個主要的好
18:36.937–18:40.500
enefits of this is that we know that research flows,
處是我們知道研究流程、
18:40.920–18:44.610
sub-agents, agent teams, all these things are super token heavy and
子代理程式、代理團隊，所有這些都非常消耗 token，
18:44.610–18:45.840
we don't really want to
我們真的不想
18:45.840–18:49.428
be chewing through our Cloud subscription on those tasks and
在這些任務上消耗我們的 Cloud 訂閱額度，
18:49.428–18:51.680
that is when we can offload those tasks
這時我們可以將這些任務
18:51.680–18:55.781
to Google servers essentially for free using Notebook LMPi and
基本上免費地轉派給 Google 伺服器，使用 Notebook LMPi，
18:55.781–18:57.560
one of the massive benefits
其中一個巨大的好處
18:57.560–18:58.581
to this is because
是因為
18:58.581–19:02.460
it's Gemini under the hood it has access to all of the YouTube content
其底層使用的是 Gemini，它可以存取所有的 YouTube 內容
19:02.460–19:03.304
transcripts and
字幕檔
19:03.304–19:04.027
all of that ric
以及
19:04.027–19:07.100
h information which you normally would not be able to surface
所有那些通常僅憑 Claude Code 無法取得的豐富資訊
19:07.100–19:09.317
just using Claude Code alone. So
所以
19:09.317–19:13.340
I asked Claude Code to use the Notebook LMPi CLI to carry out
我要求 Claude Code 使用 Notebook LMPi CLI 來執行
19:13.340–19:17.043
research on GPT 5.6 SOL versus Claude Fable 5 and
關於 GPT 5.6 SOL 與 Claude Fable 5 的研究
19:17.043–19:20.560
then turn that into an infographic. So it went
然後將其轉化為資訊圖表。所以它經歷了這裡的過程，而這是我們最終得到的結果。
19:20.560–19:23.516
through the process here and this is what we got at the end.
經歷了這裡的過程，而這是我們最終得到的結果。
19:23.516–19:25.240
Now bear in mind I did not need to do
現在請注意，我不需要做任何
19:25.240–19:29.880
anything for this. I did not use any of my clod tokens or my plan limits and
任何事。我沒有使用任何我的 clod tokens 或我的方案限制，並且
19:29.880–19:31.040
this is what we got
這就是我們最終得到的
19:31.040–19:33.252
at the end. So this is an awesome tool and
結果。所以這是一個很棒的工具，而且
19:33.252–19:36.000
you should definitely be leveraging this to save
你絕對應該利用它來節省
19:36.000–19:39.481
yourself a bunch of tokens. That is everything for this video today.
大量的 tokens。這就是今天影片的全部內容。
19:39.481–19:40.580
If you got some kind of
如果你從中獲得了某種
19:40.580–19:42.847
value out of this then please hit subscribe.
價值，請點擊訂閱。
19:42.847–19:45.420
I would massively appreciate it and make sure that
我會非常感激，並確保你
19:45.420–19:48.698
you check out the pinned comment for all of the free resources.
查看置頂評論中的所有免費資源。
19:48.698–19:50.400
You'll also find a full clod code
你還會在學校中找到完整的 clod code
19:50.400–19:54.960
masterclass in the school which will take you from complete beginner to proficient
大師課程，它將帶你從完全初學者到熟練
19:54.960–19:58.960
cloud code user so thanks for watching i'll see you in the next one
cloud code 用戶，所以感謝觀看，我們下一部影片見
影片筆記：Never Hit Claude Code Usage Limits Again
一句話總結
透過理解 Claude Code 的快取（Caching）機制與計費邏輯，並結合 clod.md、自訂技能、特定外掛（Plugins）及離線研究工具，有效將上下文使用量控制在低比例，從而避免觸發訂閱限制並大幅降低 Token 成本。
核心重點
- 快取機制是節省成本關鍵：Claude Code 不會跨回合記住用戶，每次互動都會重新發送所有歷史對話。系統識別已存在的「前綴」（Prefix）僅對新增內容收費，但快取讀取並非免費。
- 上下文使用量需嚴格控制：建議將上下文使用量控制在 20-25% 以下，以避免高昂的快取讀取費用累積及模型效能下降。
- 避免快取失效操作：
- 避免在會話中途切換模型（如 Sonnet 與 Opus 之間切換）。
- 避免在會話中途更改「努力層級」（Effort Level）。
- 保持互動在每則提示後一小時內，防止快取過期（TTL 為一小時）。
- 結構化知識管理：
- 使用
clod.md檔案自動注入專案上下文，避免重複解釋。 - 建立外部記憶系統（如 Obsidian）管理專案知識。
- 利用外掛與工具優化效率：
- 使用自訂技能（Custom Skills）打包重複任務。
- 使用 Caveman 外掛減少冗長輸出。
- 使用 Ponytail 外掛讓 AI 編碼更「懶惰」，減少程式碼行數與 Token。
- 使用 Codex 外掛分流至 OpenAI 模型進行程式碼審查。
- 使用 Notebook LMPi CLI 將重度研究任務離線至 Google 伺服器（Gemini），節省 Claude 訂閱額度。
詳細大綱
一、 Claude Code 核心運作原理與計費邏輯
- 無跨回合記憶：模型不具備跨回合記憶，每次發送新訊息時，都會將之前的所有對話內容重新發送給模型。
- 快取機制（Caching）：
- 系統識別哪些是已存在的「前綴」（Prefix）。
- 僅對新增內容收費。
- 若前綴中的任何資訊發生改變，或超過一小時未互動導致快取無效，用戶將需為整個上下文窗口重新支付全額費用。
- 計費邏輯：
- API 快取讀取費用為 0.1x。
- 訂閱方案的快取讀取比例未公開，但並非全額收費。
- 累積的快取讀取成本仍會隨回合增加。
二、 節省成本與效率的技巧（Tip 1 - Tip 6）
- Tip 1：使用
clod.md檔案
- 自動注入專案上下文。
- 避免重複解釋技術堆疊、規則與任務。
- 建立方式：使用
/init指令或描述需求。
- Tip 2：控制上下文使用量在 20-25% 以下
- 監控方式：使用狀態列或
/context指令查看視覺化數據。 - 解決方案：當接近限制時，生成「交接檔案」（Handoff file）與「會話恢復提示」（Session resume prompt）。
- 進階：使用名為
handoff的技能（Skill）。
- Tip 3：保持互動在每則提示後一小時內
- Claude Code 請求一小時 TTL（Time To Live）。
- 超過一小時未互動會使快取無效，導致下次提示需支付全額上下文費用。
- Tip 4：避免在會話中途切換模型
- 切換模型（如 Sonnet 至 Opus）會使快取失效。
- 建議策略：使用高階模型（Opus/Fable）制定計劃，然後切換至執行模型，避免中途來回切換。
- Tip 5：避免在會話中途更改努力層級（Effort Level）
- 更改努力層級同樣會使快取失效。
- Tip 6：建立適當的記憶系統
- 使用 Obsidian 等軟體建立知識圖譜。
- 結構化資料夾（工作資料夾、子資料夾）。
- 將 Obsidian 資料夾指向 Claude Code，使其能準確找到資訊。
三、 進階工具與外掛優化（Tip 7 - Tip 11）
- Tip 7：建立自訂技能（Creating Custom Skills）
- 核心價值：節省 Token、最大化個人化價值。
- 方法：將特定主題的豐富資訊打包，使其可重複使用。
- 推薦工具：Anthropic 官方技能「Skill Creator」。
- 操作流程：解釋需求 -> AI 提問與測試 -> 評估觸發條件 -> 安裝為斜線指令（Slash Command）。
- 建議：複製螢幕上的提示詞，讓 Claude 分析過去 10-20 個會話，找出重複任務並轉為技能。
- 安裝方式：執行
/plugins，在 Discover 標籤頁找到並安裝（建議選擇 User Scope）。
- Tip 8：使用 Caveman 外掛
- 目的：減少 Claude Code 冗長的輸出（Verbose）。
- 效果：保持相同答案，但輸出 Token 最多減少 65%（實際體驗可能較少，但確實有減少）。
- 原理：讓 AI 編碼代理者像「穴居人」一樣說話（簡潔有力）。
- 驗證：透過對比實驗顯示，使用 Caveman 後的回應長度顯著短於未使用時。
- Tip 9：使用 Ponytail 外掛
- 目的：讓 AI 編碼代理者盡可能「懶惰」，避免寫出不必要的程式碼。
- 效能指標（相較於無技能基準）：
- 程式碼行數減少 54%
- Token 減少 22%
- 成本減少 20%
- 時間減少 27%
- 運作邏輯（階層檢查）：
- 檢查是否真的需要新功能？
- 程式碼庫中是否已有現成解決方案？
- 能否用單行程式碼解決？
- 若以上皆否，才寫下完成任務所需的最小程式碼量。
- Tip 10：使用 Codex 外掛
- 適用對象：已訂閱 OpenAI 的使用者。
- 功能：在 Claude Code 生態系統中使用 Codex，將任務分流至 GPT 模型。
- 優勢：GPT 模型成本顯著低於 Anthropic 模型，且競爭激烈（提及 GPT 5.6 已發布）。
- 操作：安裝後執行
/codex查看指令。 - 推薦功能：Codex Adversarial Review（對抗性審查），深入檢查程式碼庫中的安全漏洞。
- Tip 11：離線重度研究任務至 Google 伺服器 (Notebook LMPi)
- 工具：Notebook LMPi CLI。
- 功能：在 Claude Code 命令列中執行深度網路研究、生成資訊圖表（Infographics）與簡報（Slide Decks）。
- 優勢：
- 研究流程、子代理者（sub-agents）通常極耗 Token，此工具可免費（利用 Google 伺服器）處理。
- 底層為 Gemini，可存取 YouTube 影片轉錄稿等豐富資訊，這是 Claude Code 單獨無法做到的。
- 案例：使用 Notebook LMPi CLI 研究 GPT 5.6 SOL 與 Claude Fable 5 的對比，並生成資訊圖表，全程未消耗 Claude 的 Token 或計劃限制。
工具 / 模型 / 名詞整理
- Claude Code：主要使用的 AI 編碼工具。
- Claude / Anthropic：模型與開發公司。
- Opus 4.8：提及的高階模型。
- Sonnet：提及的模型版本。
- Fable：提及的模型名稱（疑點）。
- Obsidian：用於建立知識圖譜的外部記憶系統。
- clod.md：用於自動注入專案上下文的檔案（疑點）。
- handoff：技能（Skill）名稱，用於生成交接檔案。
- Skill Creator：Anthropic 官方提供的技能，用於建立自訂技能。
- Caveman：外掛，用於減少冗長輸出。
- Ponytail：外掛，用於讓 AI 編碼更「懶惰」。
- Codex：外掛，用於分流任務至 OpenAI 模型。
- OpenAI：提供 GPT 模型的公司。
- GPT models / GPT 5.6：OpenAI 的模型系列（疑點）。
- Notebook LMPi：CLI 工具，用於離線重度研究。
- Google servers / Gemini：Notebook LMPi 使用的底層伺服器與模型。
- Claude Fable 5：提及的模型名稱（疑點）。
- SOL：在模型對比中出現的名詞（疑點）。
- marked-in files：描述 Obsidian 功能時出現的名詞（疑點）。
- CLODs：逐字稿中出現的名詞（疑點）。
- clod api：逐字稿中出現的名詞（疑點）。
- clod code：逐字稿中出現的名詞（疑點）。
- cloud code / CloudCode：逐字稿中出現的名詞（疑點）。
操作流程整理
- 初始化專案上下文：
- 使用
/init指令或描述需求，建立clod.md檔案。 - 將專案設定、技術堆疊、規則與任務存入該檔案，避免每次重新解釋。
- 監控與控制上下文：
- 使用狀態列或
/context指令監控上下文使用量。 - 確保上下文使用量保持在 20-25% 以下。
- 若接近限制，生成「交接檔案」（Handoff file）與「會話恢復提示」（Session resume prompt）。
- 建立與安裝自訂技能：
- 執行
/plugins進入 Discover 標籤頁。 - 安裝「Skill Creator」技能。
- 讓 Claude 分析過去 10-20 個會話，找出重複任務。
- 評估觸發條件，將重複任務打包為技能並安裝為斜線指令（建議選擇 User Scope）。
- 安裝與使用優化外掛：
- 安裝 Caveman 外掛以減少冗長輸出。
- 安裝 Ponytail 外掛以減少程式碼行數與 Token。
- 安裝 Codex 外掛，並執行
/codex進行程式碼審查（對抗性審查）。
- 執行離線研究任務：
- 使用 Notebook LMPi CLI。
- 執行深度網路研究、生成資訊圖表與簡報。
- 確認任務由 Google 伺服器（Gemini）處理，不消耗 Claude 訂閱額度。
- 維持會話穩定性：
- 避免在會話中途切換模型或更改努力層級。
- 確保每則提示後在一小時內進行互動，防止快取過期。
值得注意的限制或風險
- 快取過期風險：若超過一小時未互動，快取無效，下次提示需支付全額上下文費用。
- 上下文過滿風險：上下文使用量過高（超過 25%）會導致模型效能下降，且累積的快取讀取成本高昂。
- 模型切換風險：在會話中途切換模型（如 Sonnet 至 Opus）會使快取失效。
- 努力層級變更風險：在會話中途更改努力層級會使快取失效。
- 成本分流限制：使用 Codex 外掛分流至 OpenAI 模型需已訂閱 OpenAI，且需確認 GPT 模型的成本優勢與競爭狀況。
- 離線工具依賴：Notebook LMPi 依賴 Google 伺服器與 Gemini 模型，需確認其免費或低成本政策是否持續。
逐字稿辨識疑點
- clod.md：逐字稿中多次提及此檔案名稱，通常此類專案設定檔為
.cursorrules或.clinerules等，但逐字稿明確聽寫為clod.md，故保留原樣。 - CLODs：逐字稿中提到 "CLODs do not disclose publicly..."，疑為 "Claude" 或 "Anthropic" 的口誤或聽寫錯誤。
- clod api：逐字稿中提到 "clod api"，疑為 "Claude API" 的口誤。
- clod code：逐字稿中多次出現 "clod code"，疑為 "Claude Code" 的口誤。
- cloud code / CloudCode：逐字稿中出現 "CloudCode" 與 "Cloud Code"，疑為 "Claude Code" 的口誤。
- marked-in files：逐字稿中描述 Obsidian 功能時提到 "marked-in files"，疑為 "Markdown files" 的口誤。
- mid-session：逐字稿中多次使用 "mid-session"，意指會話中途。
- forward slash plugins：口語表達，指執行
/plugins指令。 - forward slash codex：口語表達，指執行
/codex指令。 - GPT 5.6：目前公開資訊中 GPT 版本編號通常為 3.5, 4, 4o 等，"5.6" 可能為口誤或特定內部/未來版本名稱，依規則保留原樣。
- Claude Fable 5：逐字稿中出現此模型名稱，與常見模型名稱（如 Claude 3.5 Sonnet, Opus 等）不符，可能為聽寫錯誤，依規則保留原樣。
- SOL：在 "GPT 5.6 SOL versus Claude Fable 5" 中，"SOL" 語意不明，可能為聽寫錯誤（如 Sonnet?），依規則保留原樣。
- Obsidian RAG system：逐字稿中提及此名稱，疑為 Obsidian 的 RAG（檢索增強生成）系統功能描述。
- clod.md (Obsidian 相關)：逐字稿中多次出現此名稱，結合上下文（Obsidian、專案地圖、知識庫規則），極大機率為 Obsidian.md 或特定專案名稱的聽寫錯誤，但依規則保留原樣。
可延伸追問
clod.md的具體格式與內容範例為何？- 如何精確計算上下文使用量的百分比？狀態列的視覺化數據如何解讀？
handoff技能生成的「交接檔案」與「會話恢復提示」的具體結構是什麼？- Caveman 與 Ponytail 外掛在實際編碼場景中的衝突或協作方式為何？
- Codex 外掛分流至 OpenAI 模型時，如何確保程式碼審查的準確性與安全性？
- Notebook LMPi CLI 生成的資訊圖表與簡報的品質與客製化程度如何？
- 快取過期的一小時 TTL 是否可調整？若不可調整，如何自動化維持會話活躍？
- 切換模型（如 Sonnet 至 Opus）時，是否有無縫切換的技巧或工具？
- 努力層級（Effort Level）的具體選項與對快取影響的詳細機制為何？
- 如何評估自訂技能（Custom Skills）的觸發條件以避免誤觸發或漏觸發？
生字列表
| 生字 | 讀音 | 類型 | 中文 |
|---|---|---|---|
| subscription | さくきゅう | 名詞 | 訂閱（方案） |
| overlook | オーバールート | 動詞 | 忽略、忽視 |
| interaction | インタラクション | 名詞 | 互動、交流 |
| context | コンテキスト | 名詞 | 上下文、語境 |
| caching | キャッシング | 名詞 | 快取、緩存 |
| prompt | プロンプト | 名詞 | 提示詞、提示 |
| prefix | プレフィックス | 名詞 | 前綴、前導部分 |
| documentation | ドキュメンテーション | 名詞 | 文件、說明書 |
| stack | スタック | 名詞 | 堆疊（技術棧） |
| auto-injected | オートインジェクテッド | 形容動詞 | 自動注入 |
| quota | クォータ | 名詞 | 配額、額度 |
| invalidate | インバリデイト | 動詞 | 使無效、使失效 |
| offload | オフロード | 動詞 | 轉派、卸載 |
| benchmark | ベンチマーク | 名詞 | 基準、基準測試 |
| infographic | インフォグラフィック | 名詞 | 資訊圖表 |
| leveraging | レバレッジ | 動詞 | 利用、運用 |
生字解說
subscription さくきゅう
意思：訂閱（方案）
解說：指定期付費使用軟體或服務的合約。在 AI 工具中常指付費方案。
影片原句
body who's been running into plan limits on your Claude subscription.
某個在使用 Claude 訂閱時遇到方案限制的人。
body who's been running into plan limits on your Claude subscription.
某個在使用 Claude 訂閱時遇到方案限制的人。
延伸例句
私はNetflixのサブスクリプションを更新しました。
我更新了 Netflix 的訂閱。
私はNetflixのサブスクリプションを更新しました。
我更新了 Netflix 的訂閱。
overlook オーバールート
意思：忽略、忽視
解說：沒有注意到或沒有考慮到某事。
影片原句
And there's a key fundamental principle that most people totally overlook,
而且有一個大多數人完全忽略的關鍵基本原則，
And there's a key fundamental principle that most people totally overlook,
而且有一個大多數人完全忽略的關鍵基本原則，
延伸例句
彼は重要な细节を見落とした。
他忽略了重要的細節。
彼は重要な细节を見落とした。
他忽略了重要的細節。
interaction インタラクション
意思：互動、交流
解說：雙方或多方之間的相互影響或對話。
影片原句
So if you've had 50 interactions back and forth with CloudCode in your session,
所以如果你在這個工作階段中與 CloudCode 進行了 50 次來回互動，
So if you've had 50 interactions back and forth with CloudCode in your session,
所以如果你在這個工作階段中與 CloudCode 進行了 50 次來回互動，
延伸例句
ユーザーとのインタラクションが増えた。
與用戶的互動增加了。
ユーザーとのインタラクションが増えた。
與用戶的互動增加了。
context コンテキスト
意思：上下文、語境
解說：對話或文本中出現的背景資訊或前文內容。
影片原句
it's receiving all of that context again.
在第 51 則訊息時，它會再次接收所有的這些上下文。
it's receiving all of that context again.
在第 51 則訊息時，它會再次接收所有的這些上下文。
延伸例句
文脈（コンテキスト）なしでは意味が通じない。
沒有上下文就無法理解意思。
文脈（コンテキスト）なしでは意味が通じない。
沒有上下文就無法理解意思。
caching キャッシング
意思：快取、緩存
解說：將常用數據暫時存儲以便快速訪問的技術。
影片原句
And that brings us on to the idea of caching.
這就引出了快取的概念。
And that brings us on to the idea of caching.
這就引出了快取的概念。
延伸例句
キャッシュをクリアすると、読み込みが遅くなる。
清除快取後，讀取速度會變慢。
キャッシュをクリアすると、読み込みが遅くなる。
清除快取後，讀取速度會變慢。
prompt プロンプト
意思：提示詞、提示
解說：輸入給 AI 模型的指令或問題。
影片原句
Claude is going to receive the system prompt, the project context and your initial message.
Claude 會收到系統提示詞、專案上下文以及你的初始訊息。
Claude is going to receive the system prompt, the project context and your initial message.
Claude 會收到系統提示詞、專案上下文以及你的初始訊息。
延伸例句
良いプロンプトを書くのが難しい。
寫出好的提示詞很難。
良いプロンプトを書くのが難しい。
寫出好的提示詞很難。
prefix プレフィックス
意思：前綴、前導部分
解說：在快取機制中，指對話歷史中已存在且未改變的部分。
影片原句
all of this is known as the prefix.
所有這些都被稱為前綴（prefix）。
all of this is known as the prefix.
所有這些都被稱為前綴（prefix）。
延伸例句
この単語はプレフィックスを持つ。
這個詞帶有前綴。
この単語はプレフィックスを持つ。
這個詞帶有前綴。
documentation ドキュメンテーション
意思：文件、說明書
解說：關於軟體或系統的詳細說明文件。
影片原句
CLODs do not disclose publicly in their documentation what that multiple is for planned subscriptions,
Claude 並未在公開文件中披露計劃訂閱的倍率是多少，
CLODs do not disclose publicly in their documentation what that multiple is for planned subscriptions,
Claude 並未在公開文件中披露計劃訂閱的倍率是多少，
延伸例句
ドキュメントを確認してください。
請查閱文件。
ドキュメントを確認してください。
請查閱文件。
stack スタック
意思：堆疊（技術棧）
解說：開發專案所使用的技術組合。
影片原句
you need to explain who you are, what your tech stack is,
你都需要解釋你是誰、你的技術堆疊是什麼、
you need to explain who you are, what your tech stack is,
你都需要解釋你是誰、你的技術堆疊是什麼、
延伸例句
私の技術スタックはReactとNode.jsです。
我的技術堆疊是 React 和 Node.js。
私の技術スタックはReactとNode.jsです。
我的技術堆疊是 React 和 Node.js。
auto-injected オートインジェクテッド
意思：自動注入
解說：系統自動將內容插入到指定位置。
影片原句
it gets auto-injected into your context window at every new session start in that project.
它會在該專案的每個新工作階段開始時自動注入你的上下文視窗中。
it gets auto-injected into your context window at every new session start in that project.
它會在該專案的每個新工作階段開始時自動注入你的上下文視窗中。
延伸例句
設定は自動的に注入される。
設定會自動注入。
設定は自動的に注入される。
設定會自動注入。
quota クォータ
意思：配額、額度
解說：允許使用的最大數量或限制。
影片原句
it will chew through your five hour session limit and your weekly quota on your Claude subscription.
它會消耗掉你的五小時工作階段限制以及你 Claude 訂閱的每週配額。
it will chew through your five hour session limit and your weekly quota on your Claude subscription.
它會消耗掉你的五小時工作階段限制以及你 Claude 訂閱的每週配額。
延伸例句
月の使用量クォータに達した。
達到了每月的使用量配額。
月の使用量クォータに達した。
達到了每月的使用量配額。
invalidate インバリデイト
意思：使無效、使失效
解說：使之前的狀態或快取不再有效。
影片原句
then the cache is then invalidated and your next prompt,
快取就會失效，而你的下一個提示詞，
then the cache is then invalidated and your next prompt,
快取就會失效，而你的下一個提示詞，
延伸例句
チケットは無効になった。
票據失效了。
チケットは無効になった。
票據失效了。
offload オフロード
意思：轉派、卸載
解說：將任務或負擔轉移給其他系統或人。
影片原句
which means we can offload tasks from the Anthropic models onto the GPT models.
這意味著我們可以將來自 Anthropic 模型的任務轉派給 GPT 模型。
which means we can offload tasks from the Anthropic models onto the GPT models.
這意味著我們可以將來自 Anthropic 模型的任務轉派給 GPT 模型。
延伸例句
作業を他のサーバーにオフロードする。
將作業卸載到其他伺服器。
作業を他のサーバーにオフロードする。
將作業卸載到其他伺服器。
benchmark ベンチマーク
意思：基準、基準測試
解說：用於評估性能的標準或測試。
影片原句
So if we look at the benchmarks here you can see the impact which this has.
所以如果我們查看這裡的基準測試，你可以看到它的影響。
So if we look at the benchmarks here you can see the impact which this has.
所以如果我們查看這裡的基準測試，你可以看到它的影響。
延伸例句
ベンチマークスコアが高い。
基準測試分數很高。
ベンチマークスコアが高い。
基準測試分數很高。
infographic インフォグラフィック
意思：資訊圖表
解說：將數據視覺化的圖表。
影片原句
We can also use it to generate infographics and slide decks
我們也可以用它來生成資訊圖表和投影片簡報
We can also use it to generate infographics and slide decks
我們也可以用它來生成資訊圖表和投影片簡報
延伸例句
インフォグラフィックを作成する。
製作資訊圖表。
インフォグラフィックを作成する。
製作資訊圖表。
leveraging レバレッジ
意思：利用、運用
解說：有效地使用資源或優勢。
影片原句
you should definitely be leveraging this to save yourself a bunch of tokens.
你絕對應該利用它來節省大量的 tokens。
you should definitely be leveraging this to save yourself a bunch of tokens.
你絕對應該利用它來節省大量的 tokens。
延伸例句
この機会を最大限に活用する。
充分利用這個機會。
この機会を最大限に活用する。
充分利用這個機會。
句型解說（含實例）
run into
意思：遇到、碰上（困難或限制）
接續：動詞 + into + 名詞
解說：指意外遇到問題、限制或障礙。
影片原句
body who's been running into plan limits on your Claude subscription.
某個在使用 Claude 訂閱時遇到方案限制的人。
body who's been running into plan limits on your Claude subscription.
某個在使用 Claude 訂閱時遇到方案限制的人。
實例
- 彼は予算制限に直面した。
他遇到了預算限制。 - 交通渋滞に巻き込まれた。
他碰上了交通堵塞。
back and forth
意思：來回、反覆
接續：副詞片語
解說：形容對話或動作在兩者之間來回進行。
影片原句
So if you've had 50 interactions back and forth with CloudCode in your session,
所以如果你在這個工作階段中與 CloudCode 進行了 50 次來回互動，
So if you've had 50 interactions back and forth with CloudCode in your session,
所以如果你在這個工作階段中與 CloudCode 進行了 50 次來回互動，
實例
- 彼らは議論を交わした。
他們進行了反覆的討論。 - ボールをパスし合った。
他們來回傳球。
bring us on to
意思：引出了、帶我們進入
接續：動詞 + us + on to + 名詞
解說：用於轉折，引出新話題或概念。
影片原句
And that brings us on to the idea of caching.
這就引出了快取的概念。
And that brings us on to the idea of caching.
這就引出了快取的概念。
實例
- この話題は重要な点に繋がる。
這個話題引出了一個重要的點。 - 話は次の章に進んだ。
故事進入了下一章。
chew through
意思：消耗、耗盡
接續：動詞 + through + 名詞
解說：形象地描述快速消耗資源（如 token、時間、金錢）。
影片原句
that would be chewing through millions of tokens, and it would.
這會消耗數百萬個 token，而它確實會。
that would be chewing through millions of tokens, and it would.
這會消耗數百萬個 token，而它確實會。
實例
- 彼は金を浪費した。
他耗盡了金錢。 - 時間があっという間に過ぎた。
時間被迅速消耗完了。
make sure
意思：確保、確認
接續：make sure + (that) + 句子
解說：確保某事發生或為真。
影片原句
you can make sure you never run into those usage limits again.
讓你確保永遠不再遇到這些使用量限制。
you can make sure you never run into those usage limits again.
讓你確保永遠不再遇到這些使用量限制。
實例
- 鍵をかけたことを確認した。
我確認鎖了門。 - 全員が参加していることを確認する。
確保所有人都參加了。
get into
意思：進入、深入探討
接續：動詞 + into + 名詞
解說：指開始討論或進入某個主題、狀態。
影片原句
when we get into the tips.
但現在只需理解，如果快取資料中的某些內容或前綴中的某些內容發生變化
when we get into the tips.
但現在只需理解，如果快取資料中的某些內容或前綴中的某些內容發生變化
實例
- 詳細な議論に入った。
我們進入了詳細討論。 - 問題に直面した。
他遇到了問題。
pick up
意思：繼續、接續
接續：動詞 + up + 名詞/副詞
解說：指在中斷後繼續某事，或拾起某物。
影片原句
and you can pick up exactly where you left off.
然後你就可以精確地從你離開的地方繼續進行。
and you can pick up exactly where you left off.
然後你就可以精確地從你離開的地方繼續進行。
實例
- 議論を再開した。
我們繼續之前的討論。 - 本を拾い上げた。
他撿起了書。
hands down
意思：毫無疑問地、輕鬆地
接續：副詞片語
解說：強調某事是絕對的或毫無爭議的。
影片原句
This one is hands down the most powerful tip in this video,
這一點毫無疑問是本影片中最強大的提示，
This one is hands down the most powerful tip in this video,
這一點毫無疑問是本影片中最強大的提示，
實例
- 彼は間違いなく勝者だ。
他毫無疑問是贏家。 - これは最も簡単な方法だ。
這無疑是最簡單的方法。