All-Docs Improver

Audit, score, and improve the documentation surface of a repository — CLAUDE.md, README.md, and docs/**/*.md — to keep contributors and Claude sessions oriented.

This skill writes to doc files. It always presents a quality report and waits for explicit approval before applying any edits.

Workflow

Phase 1: Discovery

Find every relevant doc file. Skip vendored / generated / build directories.

# CLAUDE.md family (handled by claude-md-improver)
find . -name "CLAUDE.md" -o -name ".claude.md" -o -name ".claude.local.md" 2>/dev/null \
  | grep -v -E "/(node_modules|\.git|venv|\.venv|dist|build|target)/" | head -50

# READMEs at any depth (skip vendored)
find . -name "README.md" 2>/dev/null \
  | grep -v -E "/(node_modules|\.git|venv|\.venv|dist|build|target)/" | head -50

# Everything under docs/
find docs -type f -name "*.md" 2>/dev/null | head -100

If none of the three sets has any files, stop and tell the user there are no docs to audit.

Phase 2: Quality Assessment

Score each file against the rubric appropriate to its type. See references/quality-criteria.md for the full rubrics.

Delegation rule: if any CLAUDE.md files were found, invoke claude-md-management:claude-md-improver via the Skill tool before assessing READMEs and docs/. Capture its quality report and merge it into the unified report in Phase 3. Do not re-score CLAUDE.md files yourself.

Cross-reference with the actual repo:

• Run (or mentally run) any documented commands; verify they exist in Makefile, package.json, pyproject.toml, etc.
• Check that referenced file paths exist (ls / Read).
• Compare claimed versions against pyproject.toml / package.json / lockfiles.
• Compare badge URLs and screenshots against current state.
• For docs/ contributor guides, spot-check that referenced commands and file paths are still valid.

Phase 3: Quality Report

Always print the unified report before any edits.

## Documentation Quality Report

### Summary
- Files audited: X (CLAUDE.md: A, README.md: B, docs/: C)
- Average score: X/100
- Files needing update: X

### CLAUDE.md family
[verbatim from claude-md-improver — do not re-score]

### README.md files

#### ./README.md
**Score: XX/100 (Grade: X)**

| Criterion | Score | Notes |
|-----------|-------|-------|
| Install / quickstart works | X/20 | ... |
| Public surface complete | X/20 | ... |
| Currency (versions / paths / commands) | X/20 | ... |
| Examples runnable | X/15 | ... |
| Conciseness | X/15 | ... |
| Onboarding clarity | X/10 | ... |

**Issues:**
- [Specific, file:line where possible]

**Recommended changes:**
- [Concrete, one line each]

### docs/

#### ./docs/architecture.md
**Score: XX/100 (Grade: X)**

| Criterion | Score | Notes |
|-----------|-------|-------|
| Accuracy vs current code | X/25 | ... |
| Structure / discoverability | X/20 | ... |
| Currency | X/20 | ... |
| Conciseness | X/15 | ... |
| Links resolve | X/10 | ... |
| Non-duplication with README/CLAUDE.md | X/10 | ... |

**Issues:** ...
**Recommended changes:** ...

Phase 4: Targeted Updates

After the report, propose concrete edits. Do not apply yet.

Update guidelines:

1. One concept per line — same density as revise-all-docs. No filler, no restating obvious code.
2. Edit existing files only — never create new doc files in this skill. If a finding doesn't fit anywhere, list it under "Unfiled findings" and let the user decide.
3. Show diffs per file, grouped. For each file that gets edits, output a ### Update: ./<path> heading, a one-line **Why:** <reason>, then

a standard ``` `diff `` fenced block with the - / + lines. Example for ./README.md where the quickstart references a removed make dev` target:

   - make dev
   + make serve

1. Surface deletions too — outdated sections, dead links, references to removed features. These often matter more than additions.
2. Do not duplicate — if the same fact belongs in both README and a docs file, pick the canonical home and link from the other.

Phase 5: Apply

Wait for explicit user approval ("yes" / "apply" / per-file selection). Then:

• Use Edit (never Write — we're amending, not replacing).
• Apply only the edits the user approved.
• After applying, summarize: files changed, lines added/removed per file, and any findings deferred or declined.

Common Issues to Flag

README.md:

• Install commands that no longer work (wrong tool, wrong path, missing step)
• Screenshots / badges referencing prior versions or deleted features
• Tool / command listings missing newly added or removed entries
• Version mismatch with pyproject.toml / package.json
• Quickstart that skips a required env-var or auth step
• Marketing prose that's drifted from what the project actually does

docs//.md:*

• Architecture diagrams referencing deleted modules
• Contributor guides citing dead make targets or scripts
• Duplicate content also covered (more accurately) in README or CLAUDE.md
• Orphan files no longer linked from anywhere
• Broken intra-doc links (](./other.md))
• "TODO" / "WIP" markers older than the file's most recent commit

What this skill does not do

• Session learnings — that's /revise-all-docs. This skill audits the static state of docs, not what was learned this session.
• CLAUDE.md scoring — delegated to claude-md-management:claude-md-improver.
• Creating new doc files — only edits existing ones. New files are a human decision.
• Code or config edits — only doc files (*.md).