Contributions are welcome.
- Prefer benchmark quality over prompt complexity.
- Keep examples public-safe and synthetic unless you have explicit permission to share real materials.
- When proposing a mutation pattern, explain what failure mode it is meant to fix.
- If you add a harness template, include at least one regression guard example.
- If you add a worked example, include both a caveat section and a reason the measured improvement is believed to be real.
Good additions:
- new harness templates
- eval pattern libraries
- worked examples for different skill types
- stronger docs on limitations and benchmark design
Avoid:
- unverifiable claims of improvement
- examples that expose private project data
- prompt mass without benchmark justification