Capture session learnings and persist them into project documentation. Run at the end of coding sessions to build a continuous learning loop.
- Reviews the session for mistakes, corrections, discoveries, and decisions
- Audits existing docs to avoid duplicates and find the right placement
- Prunes stale entries when CLAUDE.md is above 250 lines (duplicated by docs, enforced by toolchain, outdated, or session debris)
- Proposes targeted updates with progressive disclosure (pithy CLAUDE.md + detailed .claude/rules/)
- Applies with approval — nothing is written without your say-so
"Debrief this session"
"Capture learnings from what we just did"
"Update project knowledge with what we learned"
"What did we learn? Save it."
| Type | Destination |
|---|---|
| Universal rules | CLAUDE.md (one-liners only) |
| Domain-specific knowledge | .claude/rules/*.md (path-scoped) |
| Architectural decisions | ARCHITECTURE.md or ADRs |
| Process changes | CONTRIBUTING.md |
The skill enforces token-efficient documentation:
- CLAUDE.md stays under 300 lines
- Each addition must pass: "Would removing this cause mistakes?"
- Detailed knowledge routes to .claude/rules/ with path-scoping
- Prefers @imports over inline content
Bash (git read-only), Read, Edit, Write, Glob, Grep, AskUserQuestion
- Git repository (for context about recent changes)
- EXAMPLES.md — Good vs bad debrief examples, routing decisions
- SKILL.md — Full workflow loaded by Claude Code