Skip to content

Lucineer/i-know-kung-fu-research

 
 

Repository files navigation

I Know Kung Fu: Claude Code Research

Sometimes you add a small file to your project and Claude Code's behavior changes. This tool helps you observe and test those changes in a controlled environment.

✅ Live test bed: https://the-fleet.casey-digennaro.workers.dev

Quick Start

  1. Fork this repo first – you run your own isolated tests. No account required.
  2. Deploy directly to Cloudflare Workers. Zero dependencies, zero build steps.
  3. Edit one file in /skills/. Run the test. Watch how Claude's output shifts.

You can observe changes in token ordering, context loading, and rule activation. No black box.

Why This Exists

When using Claude Code, you might notice that adding a small file sometimes improves performance. Official documentation doesn't explain this. We built this to test and document what actually happens, with reproducible results you can verify yourself.

Features

  • Trace which lines Claude prioritizes when scanning your project
  • Run side-by-side tests to measure behavior changes per line added to a Skill
  • Verify which rules persist between chat sessions
  • Observe when Claude delegates work to a subagent and what triggers it
  • Inject controlled failures to map system limits
  • Tests run against real production wifi driver code, not trivial examples
  • Every observation is marked with Claude version for tracking changes

What Makes This Different

  1. No hype – this isn't a product. We share test results.
  2. Test-driven – every claim has a runnable test you can execute.
  3. Transparent – we document how the system works, not "secrets" or "hacks."

Limitations

One specific limitation: the test bed currently only supports one codebase (a production wifi driver) for consistency. You cannot test against your own arbitrary codebase without modifying the project.

License

MIT. Use, break, and improve this however you want.

Research by Superinstance and Lucineer (DiGennaro et al.)

Releases

Packages

Contributors

Languages