skills/audit-skill-craft/SKILL.md
Judge a SKILL.md against craft axes (single responsibility, description distinctiveness, imperative voice, verifiable completion, calibration, progressive disclosure) and apply the fixes the audit surfaces. Do not use for format-presence-only checks (use reviewer-prompt) or reproducibility loops (use /tuning).
npx skillsauth add thkt/claude-config audit-skill-craftInstall this skill globally with one command. Works with Claude Code, Cursor, and Windsurf.
3 of 9 scanners reported clean
Some scanners were skipped, did not run, or reported a non-clean status. Review each row below.
Judge a SKILL.md against the craft axes in ~/.claude/rules/conventions/SKILLS.md and the format rules in ~/.claude/rules/conventions/MARKDOWN.md, then apply the fixes the audit surfaces. This pairs the format-presence judgment of reviewer-prompt with craft-quality judgment and in-place repair.
Four tools touch prompt files. This table draws the boundary so the audit avoids duplicating them.
| Tool | Scope | Edits files | | ------------------------ | ------------------------------------------------------------- | -------------------- | | audit-skill-craft (this) | Craft quality of a SKILL.md plus the format fixes it surfaces | Yes (Edit) | | reviewer-prompt | Format presence (bold, frontmatter fields, prose-to-table) | No (read-only) | | tuning | Empirical reproducibility loop via fresh subagents | Yes (one patch/iter) | | polish | General AI-slop cleanup across any file | Yes |
$ARGUMENTS may carry a skill path and optional flags. Resolve before use.
skills/*/SKILL.md in git status, then by mtime--no-format skips the reviewer-prompt spawn in Step 4 for a craft-only pass| Step | Action |
| ---- | --------------------------------------------------------------------------------- |
| 1 | Resolve the target SKILL.md (auto-detect or $ARGUMENTS) |
| 2 | Load the judgment basis: SKILLS.md Craft axes + MARKDOWN.md format rules |
| 3 | Judge the target against the 6 craft axes; record each verdict with line evidence |
| 4 | Spawn reviewer-prompt via Task for format findings (skip with --no-format) |
| 5 | Merge craft + format findings; classify each fix as mechanical or judgment |
| 6 | Apply mechanical fixes via Edit; present judgment fixes for confirmation |
| 7 | Re-read the file; verify it now passes the axes it failed and adds no violation |
| 8 | Print a summary table of axes, verdicts, and applied fixes |
Judge against the 6 axes in the SKILLS.md Craft section. For each axis, record pass or fail plus the line(s) that triggered the verdict.
Mechanical fixes preserve meaning; judgment fixes change wording or structure. Surface judgment fixes as proposed text and apply only the confirmed ones.
| Class | Examples | Handling | | ---------- | ---------------------------------------------------------------------- | ---------------------------- | | Mechanical | bold to table, pseudo-heading to heading, trailing-summary cut, filler | Apply via Edit | | Judgment | description rewrite, responsibility split, adding a stop condition | Propose; apply after confirm |
When an axis flags 2+ unrelated responsibilities, do not merge or split silently. Present the split boundary and ask via AskUserQuestion whether to split into sibling skills or keep as-is with a recorded exception.
Re-read the edited file. Confirm every previously-failing axis now passes and no fix introduced a new MARKDOWN.md violation. A fix that trades one violation for another is not done.
documentation
Generates and updates .claude/OUTCOME.md interactively. When the file is absent or empty (no Behavior / all sections TBD), collects content via AskUserQuestion and writes the stub; when present, shows the current state and applies updates.
tools
Internal helper for /think Step 11. Renders SOW.md + Spec.md as an integrated Astro view and returns a dev server URL.
development
Extract repository spec while detecting bugs, spec gaps, and consistency drift via dual-purpose documentation. OUTCOME.md-axis question-driven exploration with ephemeral output. Do NOT use for code review (use /audit or /polish), feature implementation (use /code), planning only (use /think), or single-bug fix (use /fix).
development
Discover undocumented design decisions and challenge each candidate via critic-design before promotion. Rank by impact and reversibility, produce ADR promotion candidates. Treat each candidate as a position arguing for ADR status, not a fact to be filed. Pairs with audit-adr-drift, which scans existing ADRs for drift against code.