Resources Prompt
PromptContext window diagnosis prompt
Paste it into the agent you want to check. It reads what is loaded globally, what the project adds, and what the session has picked up, then ranks what is actually worth your attention.
One prompt, a ranked diagnosis.
Setups rot quietly. A rule in the global config contradicts one in the project. Two tools do the same job. A file gets loaded on every single turn and gets read about once a month. Nothing breaks loudly, so nothing gets fixed.
This prompt makes the agent inspect its own setup and report on it. It looks at the global layer, the project layer, the tools, the skills, and the session history, then sorts what it found by what is worth doing first.
Everything the prompt checks.
It inspects broadly, then ranks findings by impact, confidence, frequency, effort, risk.
- Global instructions: user-level config and memory
- Project instructions: repo rules and CLAUDE.md
- Connected tools and integrations
- Available skills, commands, agents
- The session history itself
- Three buckets out: do now, investigate, low priority
- Five answers per finding: problem, action, benefit, risk, verification
In order.
Run it inside the agent you want to check, not a fresh one. It can only report on what it can see.
- Open the agent you actually work in, with its usual config loaded.
- Paste the prompt. Let it finish before you reply.
- Work the DO NOW list first. Ignore the rest until that list is empty.
- Re-run it after the changes. The list should be shorter.
Take it now.
One click copies the whole block. Paste it into Claude, ChatGPT, or a Claude Code session.
You are auditing the setup of the agent you are running inside.
Inspect everything currently loaded, in this order, and say what you
actually found rather than what you expect to find:
1. Global instructions (user-level config and memory)
2. Project instructions (repo-level config, CLAUDE.md, rules files)
3. Tools and integrations that are connected right now
4. Skills, commands and agents available to you
5. The conversation history in this session
Then list every problem you found: contradictions between layers,
instructions that never fire, tools that duplicate each other, files
loaded on every turn that are rarely used, and anything ambiguous
enough that you would have to guess.
Rank the findings into three groups:
DO NOW highest impact, high confidence
INVESTIGATE likely a problem, needs checking first
LOW PRIORITY real, but it can wait
Rank by impact, confidence, how often it bites, effort to fix, then
risk of fixing it.
For each finding, answer these five, in this order, one line each:
1. What's wrong? the problem you found
2. What should I do? the exact change you recommend
3. What do I gain? why the change is worth making
4. What could go wrong? what to check before changing anything
5. How do I check? the simplest way to see if it helped
Do not change any files. Report only.Build it with other founders.
The community opens soon. Weekly builds, the full library, and answers when this file does not match your business.