This video is a little different
It is a real debugging story from practice
I asked an AI agent to make a daily AI trend analysis for me
During the run
it had actually obtained the correct data
but in the end it made up a fake report for me
What is even more interesting is
after I fixed this problem
and the system went live
I fell into another trap
You may run into both of these traps too
Let me start with the background
As an AI channel creator
I need to track hot topics in AI every day
to choose video ideas
So I wanted an AI agent every early morning
to collect data for me automatically
from three sources
popular AI posts on X
AI projects on GitHub Trending
and related trend movements on Google Trends
When I open my computer in the morning
I can see the topic report for the day
I currently run two AI agents in parallel
OpenClaw and Hermes agent
On one hand I use them for topic selection
and on the other hand I test their capabilities
On the Hermes agent side, by building its own skill
it can already complete this report fairly well
But OpenClaw had a problem
because of fake IP restrictions
and Web fetch restrictions
it could not get the latest X data
I did not want to use the X API for now
so I decided to try another approach
Let OpenClaw
directly control a Chrome browser already logged in to X
and use the browser to collect the data
Now let us start the setup
Step one, enable Chrome remote debugging
Step two, modify the OpenClaw config file
Configure a profile called xbrowser
specifically for this project
and set the port properly
Step three
use a command to start Chrome in debug mode
Step four, log in to your X account in this Chrome
Finally, verify it with a command
to see whether OpenClaw can take over this page
From Chrome we can see
it opened a page that is already logged in
The takeover worked
There is one trap here to be aware of
Starting from Chrome 136
remote debugging port and remote debugging pipe
can no longer attach to the default user directory
They must be used with a non default user data dir
This is a change Chrome made for security reasons
So here
I created a dedicated profile called xbrowser
After setup, I ran the first test
and sent the prompt to OpenClaw
The result came back quickly
It looked pretty good
Wait, something is wrong
Look at the GitHub Trending section
Why are the recent popular projects missing
For example Hermes agent
And why are there crypto projects
I opened the original GitHub Trending page
to compare it
Number one is Andrej Karpathy skills
Number two is Claudemem
But these projects are not in the OpenClaw report at all
What is going on
Let us investigate step by step
First, look at OpenClaw
and see what raw data it got from the browser
With these two commands, we can see
the raw snapshot output is in snapshot
Andrej Karpathy skills and Claude mem are both there
So the data is correct
Was OpenClaw lying
So I added stronger constraints to the prompt
requiring it to use only the snapshot data
But the result still failed
It still used data from who knows where
It looked like the LLM was lying
In other words, the MiniMax M2.7 I was using was lying
I asked Claude separately about this phenomenon
and it gave me an explanation
It said this was context neglect
combined with hallucination from prior
context neglect means
the model received new data but ignored it
hallucination from prior means
the model instead outputs
old content it remembered from training data
The theory is too dry
Let us explain it with three analogies
First, reciting is easier than reading
It is like during an exam
there is reference material next to the question
but the student is too nervous
and directly recites what they studied before
Projects like system-design-primer have dominated Trending for a long time
and appeared countless times in training data
When the model sees GitHub Trending
it automatically associates it with these familiar names
Second, weak instruction following
Use only snapshot data
This instruction
requires the model to actively suppress its prior knowledge
But small models and distilled models
cannot do this kind of precise control
MiniMax is exactly this kind of model
Third, attention decay in long context
The snapshot data is very long
The model may not have read it carefully
and lazily used the default answer
What should we do
Try a smarter model
So I switched the model from MiniMax M2.7 to GPT 5.4
with the same prompt
and the same snapshot data
GPT 5.4 produced a completely correct result
consistent with the current GitHub Trending page
You can clearly feel that
for long text plus strong constraint tasks like this
models such as ChatGPT and Claude
and Gemini
are indeed better at reading context
and suppressing prior assumptions
by a noticeable level
So how do we prevent this kind of issue
First, add protection at the prompt level
Make rules and constraints as specific as possible
Second, use a stronger model for critical tasks
for example GPT 5.4
Third, and most importantly
never blindly trust AI output
just like every AI itself reminds you
AI can make mistakes
Always check the results
After switching the model to GPT 5.4
I added this into the OpenClaw scheduled task
and the system officially went live
But the next day, when I opened the report, there was another problem
The system reported a timeout error and did not get X data
But I saw that the background Chrome X page was actually open
My first reaction was to keep adjusting the prompt
and keep changing the OpenClaw system config
But the more I tuned, the messier it became
until I remembered the control group, Hermes agent
Hermes had the same task
and had been steadily outputting AI trend reports
It was still using MiniMax M2.7
and had not switched to GPT 5.4
At that moment I realized
this problem might not be in the model
but in the execution method
Did Hermes agent have a better method
I found the skill generated by Hermes itself
and asked Claude to break down what it was doing
Claude found that Hermes bypassed the snapshot path
It directly used Python plus WebSockets to connect to CDP
which is Chrome DevTools Protocol
Then it executed JavaScript to extract tweet data directly
Fast, stable, and not dependent on browser operations
The key point is
this method can be migrated to OpenClaw completely
I asked Claude to write a set
of Python programs for OpenClaw
with a new matching prompt
The updated process is roughly like this
X data is collected by a Python program
which interacts with Chrome through CDP
GitHub Trending
is collected directly with HTTP requests plus regular expressions
Google Trends is collected with the pytrends library
In this way
at the data source layer
we no longer depend on expensive browser based
automation
The model is only responsible for summarizing and analyzing the data
Stability improved
and execution speed became much faster
A 10x improvement is not exaggerated
The task went live again
and has been running stably for several days
The result is also very satisfying
This task can pause here for now
I will keep observing it later
Looking back
I have three thoughts from this hands on test
First
AI agents really can directly control your browser
but this method is unstable
and it also slows things down
Second, AI can make mistakes
especially small models
and models trained with distillation techniques
Do not save money here for critical tasks
Third, and the most interesting point
when one AI agent cannot do something well
you can let it learn from another AI agent approach
Other stones can polish jade
This old saying also applies in the AI field
That is it for this episode
If this episode helped you
please like and subscribe
In the next episode, I will continue sharing more AI practice experience
See you next time