What problem it solves

You know what Claude contributed at the start of a session. Two hours later, the boundary blurs. Claude's suggestions, your decisions, and the moments where AI quietly shifted direction — they evaporate.

collab-proof is an assisted retrospective that runs after a Claude Code session and keeps a record before it's gone. Think of it as an ESLint for AI collaboration — not something that blocks or measures precisely, but something that catches the moments where AI shifted direction and records them.

What's included

• collab-proof/SKILL.md — full skill with 4-layer pipeline (signal detection → intent classification → token analysis → output generation)

The core output

The AI contribution field — honest, not promotional:

``` AI contribution:

• Identified: TOCTOU window between ZCARD and ZADD developer had not noticed
• Suggested: Lua EVAL approach after reviewing Redis atomicity guarantees
• Developer-driven: Final implementation, choice of Lua over MULTI/EXEC ```

vs sessions where Claude just executed instructions: ```

• Developer-driven session. Claude executed instructions. ```

The rubric forces both sides to be named. Read the output and correct it if wrong.

Notes

• Zero external dependencies. No pip install required.
• ./install.sh single-command install
• SessionEnd and Stop hooks run async — Claude Code is never blocked
• D score is a directional indicator, not a precise metric