If you use ChatGPT or Claude Code, everyone should do this.
Even if you are using the latest models like Astra or Fable, without doing this:
**・You only get about half of their true performance.
・You hit usage limits immediately.**
"Did the AI just get dumber?!"
You might be wasting potential by using it incorrectly.
So today I will introduce
5 reviews to unlock the true performance of cutting-edge models.
Please make sure to do these.
※Copy-paste this article into an AI and ask "Inspect my AI setup based on this article," and it will handle it nicely.
OpenAI has also officially released guidance, but....
Old AGENTS.md files and Skills need maintenance to match the latest models.
Instructions written for old models are holding back current AI.
That is why you need to clean up once.
I have attached short prompts you can paste directly for each review step. At the end, there is a prompt that inspects everything at once.
Let's begin.
Why does "The AI got dumber!?" happen?
Until now, you likely added instructions to make AI work as intended.
Skills, AGENTS.md (rulebooks read every time), and prompts with "Always do ~".
Back then, that was correct.
But Astra has become smarter; it works without needing so much detail.
Yet, the instructions remain and are loaded every time you work.
It's like reading all company rules just to fix one character in a banner.
No wonder usage drops quickly lol
And this is invisible. So even if it's slow or hits limits, people blame the model being heavy.
The cause isn't the model, but the baggage we hand over.
There are 5 things to review across 3 areas: Skills, AGENTS.md, and Prompts.
1. Reviewing Skills
Skill names and descriptions are always in the AI's head to decide when to use them.
This means parts are read every time. There are 3 problems here.
OpenAI Official
① Descriptions are too long or too broad
If you add too many skills, descriptions get auto-truncated. Then the AI doesn't know which skill to use.
Worse is when activation conditions are too broad.
For example, a skill labeled "Use for design-related tasks." This gets loaded even for resizing one image.
"Use when creating or modifying the first view of an LP."
Narrow it down to this. Short, but clear on "when to use." This is what official guidelines call a good description.
② Writing everything in one file
Reading skills consumes usage.
Official recommendation: Make the first page of a skill a "Table of Contents." Write minimal info at the entry point, split detailed steps/materials into separate files. Let the AI fetch only what's needed.
Stop handing over thick manuals every time.
③ Writing steps too granularly
This was surprising.
Previously, detailed itineraries yielded better results. But current models understand nuance and ambiguity, so overly detailed instructions worsen results, according to official docs.
Carefully written manuals restrict AI judgment. The era where adding more improved results is over.
Paste this to inspect skills:
1Inspect all skills in this project.2Check for 3 points:3- Descriptions too long or activation conditions too broad4- Everything in one file; should be split into TOC and materials5- Steps too granular; unnecessary for current models6Do not rewrite yet. List "Where, Why, How to fix" for each skill.
#
2. Reviewing AGENTS.md
AGENTS.md is read every time you work in that location. It's worth reviewing line-by-line: "Is this still necessary?"
Claude Code users: Treat CLAUDE.md similarly.
OpenAI Official
④ Remove "Read every time," "Always test," "Always confirm"
Official bad examples include: "Before editing, always read these 3 documents."
Making it read 3 design docs to fix a typo burns usage and slows work.
Good example:
"For overall structure, see A; for data changes, see B; for publishing prep, see C."
Pointing to docs is OK, but contextually. Astra decides what to read.
"Always test" and "Always confirm" are similar. Old models skipped verification unless told. Astra verifies proactively.
Leaving these instructions causes redundant checks. Wasted usage.
Also, strong prohibitions.
Many wrote "Must get permission before proceeding!!" because old models went rogue.
I did too.
Astra has better judgment; it won't act if unsure. Strong prohibitions make it stop unnecessarily.
Stop → Confirm → Resume. Each cycle wastes usage and time.
Official advice: Write permissions, not prohibitions.
"This test doesn't affect production, so run it without confirmation, and fix if it fails."
Define safe zones upfront. This reduces stops significantly.
1Inspect AGENTS.md (or CLAUDE.md) line-by-line.2Check for 3 points:3- Instructions forcing doc reads regardless of task relevance4- Commands like "Always test/confirm" that current models handle automatically5- Overly strong prohibitions that halt workflow unnecessarily6Do not rewrite yet. List "Keep/Delete/Change to Permission" per line with reasons.
#
3. Reviewing Prompts
Finally, how you ask. This had the biggest new insights.
⑤ AI stopping mid-task is because you asked it to
Since switching to Astra, "Please confirm" responses increased. Many feel this way.
I initially thought it became cautious and hard to use.
But official docs explain differently. Astra judges carefully where to stop.
If your prompt lacks a "completion state," it stops at the safest point: "After creating one version." Even if more work remains.
It's not lazy; it's considerate.
Each stop requires you to say "Continue," and the AI re-reads context. This wastes usage and time.
Solution: Define the "completion state" upfront.
"Create, run, check results, and fix issues until done." If you say this, it won't return midway.
If you want broader exploration, specify what to explore and where to stop.
Conversely, if you habitually write "Give me ideas first. I'll confirm before proceeding," be careful. Only write this if you truly need confirmation. Otherwise, delete it. Astra follows vague checkpoints literally.
Vague requests lead to vague stops. Prompts should define the "finish line position," not micro-steps.
Inspect frequent prompts with this:
1I will paste my common prompts.2Inspect for 3 points and provide revised versions:3- Is the completion state defined? (Create/Run/Check/Fix)4- Are there unnecessary confirmation checkpoints?5- Are steps too granular, removing judgment room?6Add 1-line reason for each change in the revised version.7(Paste your prompt here)
Prompt to Inspect Everything at Once
Thought "Reviewing 3 areas is tedious"?
Don't worry. Official docs say: "You don't need to manually review everything. Let Astra audit it."
Paste this to inspect 1-3 at once.
▼ Copy-Paste Prompt (Start)
Read this official article and audit my project's Skills, AGENTS.md (or CLAUDE.md), and common prompts based on its content.
Article: https://developers.openai.com/blog/rethinking-skills-and-prompts-for-gpt-6-astra
Check for:
- Skills with long/broad descriptions
- Skills needing TOC/material separation
- Skills with overly granular steps
- AGENTS.md with forced reads/tests/prohibitions
- Prompts lacking completion states or having unnecessary checks
Do not rewrite yet. List fixes by impact order. I will choose what to apply.
▲ End
Key tip: Don't let it rewrite immediately; list options first.
Humans decide which rules to keep vs. delete. Leaving it entirely to AI might erase important preferences.
#
Summary
Today's 5 points:
**① Keep skill descriptions short/narrow
② Use TOC for skill first pages
③ Avoid overly granular steps
④ Remove "Always read/test/confirm"; prefer permissions over prohibitions
⑤ Define completion state in prompts**
Common theme:
If AI got smarter, reduce instructions.
We focused on "adding" for new models. Current AI needs "subtracting."
Same price/model, but lighter baggage yields better results and usage efficiency.
Think AI got dumb? Clean your environment before paying for premium models.
New model? First, deep-clean your instruction sets.
This is crucial.
Too busy? Copy-paste this article and ask AI to "Inspect my environment based on this."
Official article here 👇
https://developers.openai.com/blog/rethinking-skills-and-prompts-for-gpt-6-astra
Update.
Made something amazing. Lol
Many said: "I pasted it but don't know what to fix."
So I made it a Skill.
'Instruction Deep-Clean Skill'
Just ask "Deep-clean my instructions," and it handles:
Diagnosis → Proposal List → Selected Fixes → Report
It diagnoses:
- How much baggage your AI carries
- Which skills weren't called in 30 days
- Cost/time savings from removal
Lists high-impact fixes, applies selected ones, and generates a report.
Original files backed up automatically. No unauthorized rewrites.
