What problem does it solve?
Skills executed within Claude Code can accumulate hidden friction, bloat token consumption, run on misaligned models, or produce incorrect outputs without any systematic feedback mechanism. This Skill closes that gap by auditing a skill's actual execution trace from its session transcript, surfacing performance anomalies, permission gaps, and accuracy issues that remain invisible during the run itself.
Core Features & Use Cases
- Four-Dimension Evaluation: Judges runs across friction findings, performance metrics (wall-clock time, token usage, cache-hit ratio, cost estimate), model-fit profile, and output accuracy.
- Mechanical Extraction + Human Judgment: Runs a deterministic Node.js extractor over JSONL transcripts to produce structured findings, then classifies them with judgment to propose concrete, root-cause-targeted edits.
- Cross-Session Intelligence: With --last N or --all, aggregates recurring-friction tallies, median/p90 timing, and total cost across sessions to distinguish one-off noise from genuine drift patterns.
- Anti-Pattern Naming: Every proposed fix names the specific failure shape (bash invocation, tool sequence, or phrasing) so future agents recognize and skip the dead end.
Quick Start
Use the audit-skill skill to evaluate the most recent execution of the pr skill by invoking the slash command /audit-skill pr and review the surfaced findings and edit proposals.
