AI agent self-audit skill for Claude Code.
Checks project documentation freshness, finds recurring errors in git history, generates guardrails, and produces actionable fixes — so the agent stays calibrated to the real state of your project.
| Block | Description |
|---|---|
| Documentation & Context | Validates CLAUDE.md, ARCHITECTURE.md, ROADMAP.md, CHANGELOG.md against the actual codebase |
| Recurring Errors & Guardrails | Analyzes git history for fix/revert patterns, generates rules and automated checks |
| Proactivity | Surfaces next steps from ROADMAP, TODO/FIXME trends, and high-leverage actions |
| Confidence Calibration | Identifies blind spots — parts of the project the agent understands poorly |
| Connections & Insights | Finds logic duplication, implicit dependencies, priority mismatches |
- Full — all 5 blocks + report + auto-apply. After milestones or major refactors.
- Quick — blocks 1-2 (docs + errors). After a productive session.
- Targeted — single block on demand.
Use /agent-audit or say:
- "audit", "self-check", "meta-review"
- "check CLAUDE.md", "what's outdated", "update context"
- "what errors do you keep making", "what should we lock in"
- "we finished the feature", "sprint done", "what's next" (after major work)
Does not trigger on small fixes or cosmetic commits.
Copy agent-audit.md to your Claude Code commands directory:
# Global (available in all projects)
cp agent-audit.md ~/.claude/commands/
# Per-project
cp agent-audit.md .claude/commands/Generates docs/audit-report-YYYY-MM-DD.md with:
- Prioritized findings (critical / warning / ok)
- Concrete file edits to apply
- Questions to close knowledge gaps
- Suggestions for next steps
MIT