skill-doctor

Run the skill and choose which conversations to grade: the current repository, selected projects, or all local conversations. It samples recent sessions, scores each transcript on efficiency and code quality, and produces a letter grade with findings.

The collector supports Claude Code project history, Codex rollout session files, and Warp's local conversation databases. The default auto mode scans every locally available source, and flags allow narrowing to a single harness.

No. Everything runs locally; transcripts, session files, and excerpts are never uploaded. The only shareable artifact is the HTML report the user explicitly chooses to post.

Each sampled transcript is scored against rubrics with labeled numeric values. Raw means are curved with 0.5 + 0.5 * score, and the overall grade weights efficiency at 0.5, code quality at 0.35, and skill coverage at 0.15.

If the sampled session count is zero, the run stops and reports that there is nothing recent to score in the selected scope. It suggests raising the lookback window with the days flag or choosing different projects.

No. Proposed improvements are written to a temporary report directory as full revised SKILL.md files with unified diffs. Real skill files are only changed if the user later approves applying the suggestions.