docs-consistency

name	docs-consistency
description	Review user-facing documentation for accuracy, consistency, and completeness across private, public, nested repos, and the user manual. Use when docs feel stale, after major changes, or before sharing. (Replaces `repo-doc-audit`)
allowed-tools	Read, Glob, Grep, Bash(ls*), Bash(wc*), Bash(find*), Bash(git log*), Bash(diff*), Bash(cat*), Write
argument-hint	[scope] — all (default), private, public, cross-check, user-manual

Docs Review

Cross-cutting documentation review that checks consistency, accuracy, and completeness across the entire documentation ecosystem. It verifies that docs agree with each other and with reality; installed per-repository audit or sync workflows can deepen the check but are not required.

When to Use

• After adding, removing, or renaming skills, hooks, agents, or rules
• After major infrastructure changes (consolidation, new repos, renamed components)
• Before publishing or sharing the public repo
• Periodic health check (monthly)
• When the user says "review docs", "check my docs", "are my docs consistent"

When NOT to Use

• For per-repo prose quality — use an installed repository-documentation audit if available
• For mechanical count/freshness sync — use the repository's documented sync command
• For code architecture — use an installed external/code audit workflow

Argument Parsing

Parse $ARGUMENTS for scope (optional, default all):

Scope	What it covers
all	All 5 checks below
private	Checks 1-3 on private docs only
public	Checks 1-3 on public docs only
cross-check	Check 4 only (public-private consistency)
user-manual	Check 5 only (user manual alignment)

Checks

Check 1: Count Consistency

Count actual components on disk, then verify every document that states counts matches reality.

Disk counts (source of truth):

Skills:  find skills -name SKILL.md -type f | wc -l   # recurse — skills/{engineering,research}/ are nested; ls -d skills/*/SKILL.md MISSES them (undercounts by ~5)
Hooks:   ls hooks/*.sh hooks/*.py hooks/*.mjs 2>/dev/null | wc -l   # hooks are .sh + .py + .mjs, NOT just .sh (counting .sh-only undercounts by ~15)
Agents:  ls .claude/agents/*.md | wc -l
Rules:   ls rules/*.md | wc -l

Counting gotchas (these caused false-positive drift reports on 2026-06-06): skills live in nested category dirs too (skills/engineering/, skills/research/) so you MUST find -name SKILL.md, not glob skills/*/; and hooks are written in shell, Python, and Node, so count all three extensions.

Documents to verify (private scope):

File	What to check
README.md	Skill category table totals, hook count in heading, agent count, rule count
CLAUDE.md	Skill count, hook count, agent count, rule count, file structure table
docs/components/skills.md	Total count in header, per-category counts, overview table row count
docs/components/hooks.md	Total count in header, hook table row count
docs/components/agents.md	Total count in header, agent table row count
docs/components/rules.md	Total count in header, rule table row count
docs/system.md	Component counts in overview section

Documents to verify (public scope):

File	What to check
public/public-repo/README.md	Counts in auto-generated marker sections
public/public-repo/CLAUDE.md	Skill/hook/agent/rule counts

Report format per file:

| File | Component | Stated | Actual | Status |
|------|-----------|--------|--------|--------|
| README.md | Skills | 129 | 130 | MISMATCH |
| README.md | Hooks | 21 | 21 | OK |

Check 2: Component Coverage

Every component on disk must appear in its catalogue file. Every entry in the catalogue must exist on disk.

Component	Catalogue file	Disk location
Skills	docs/components/skills.md	find skills -name SKILL.md (incl. nested categories)
Hooks	docs/components/hooks.md	hooks/*.{sh,py,mjs}
Agents	docs/components/agents.md	.claude/agents/*.md
Rules	docs/components/rules.md	rules/*.md

For each: list items on disk not in catalogue (undocumented) and items in catalogue not on disk (orphan entries).

Also check README.md skill category table — every skill on disk should appear in exactly one category row.

Check 3: Stale References

Grep documentation files for references to components that no longer exist on disk.

Scan these files:

• README.md, CLAUDE.md
• docs/*.md (all)
• the user-manual .tex source (in docs/reference/user-manual)
• skills/shared/*.md
• rules/*.md

What to look for:

• Skill names (/skill-name or backtick-quoted) that don't match any skill on disk — resolve via find skills -name SKILL.md (a name like skill-latex is the MCP-tool form of the latex skill, NOT a separate/orphan skill — strip the skill- prefix before checking)
• Hook script names that don't match any hooks/*.{sh,py,mjs} file
• Agent names that don't match any .claude/agents/*.md file
• Broken relative links (markdown [text](path) where path doesn't exist)

Exclude log/ and MEMORY.md — these are historical records, not active documentation.

Check 4: Public-Private Cross-Consistency

Verify that public and private docs agree on the same facts.

1. Count alignment: Compare counts in public/public-repo/README.md vs README.md. They should match (sync-to-public.sh propagates counts).

2. Section currency: For each synced section (Architecture, Workflows, Session Continuity, Design Principles, Credits), check that the public version matches what extract_section() would produce from the private README. Flag sections where content has diverged.

3. Marker freshness: Check auto-generated markers in public README (COMPONENT-TABLE, SKILLS-SUMMARY, AGENTS-TABLE, HOOKS-TABLE, RULES-TABLE, FILE-TREE, FILE-STRUCTURE). Run uv run python scripts/generate-public-docs.py --dry-run if available, or manually compare marker content against disk reality.

4. Leaked details: Grep public repo for personal details that should have been anonymised:

   • Institutional names (your affiliations) outside of credits/attribution
   • Personal names (other than in credits)
   • vault file paths
   • GitHub usernames (other than user in attribution)

Check 5: User Manual Alignment

Check the user-manual .tex source (in docs/reference/user-manual) against current system state.

1. Skill tables: Extract skill names from LaTeX tabular environments. Compare against actual skills on disk.
2. Hook tables: Same for hooks.
3. Agent/rule counts: Check any stated counts.
4. Category structure: Verify skill categories in the manual match the categories in docs/components/skills.md.
5. Architecture diagram: If the manual contains a system diagram, check that component names match current naming.

Check 6: Package Coverage

Every packages/<name>/ directory (each is a package — nested git repo or local) must appear in docs/components/packages.md, and the stated package count must match disk.

When the repository declares a deterministic inventory checker, run it first. For example:

uv run python scripts/check_inventory.py --check

• Discover the actual command from project guidance, pyproject.toml, or the repository's scripts; do not assume the example filename exists.
• If there is no checker, compare package directories, documented rows, and stated counts directly.
• Count drift: stated package counts must equal the on-disk count.
• Coverage drift: every package directory must have a catalogue row, and every catalogue row must resolve to a package or an explicit external item.

Any deterministic checker failure is FAIL (or WARN for 1–3 documentation items when the repository's policy permits that severity).

Output

Write report to log/audits/docs-consistency-YYYY-MM-DD.md:

# Documentation Review — YYYY-MM-DD

## Scope: [all / private / public / cross-check / user-manual]

## Dashboard

| Check | Status | Issues |
|-------|--------|--------|
| Count Consistency | OK/WARN/FAIL | N |
| Component Coverage | OK/WARN/FAIL | N |
| Stale References | OK/WARN/FAIL | N |
| Public-Private Sync | OK/WARN/FAIL | N |
| User Manual | OK/WARN/FAIL | N |
| Package Coverage | OK/WARN/FAIL | N |

## Count Consistency
[Per-file table from Check 1]

## Component Coverage
### Undocumented (on disk, not in catalogue)
### Orphan Entries (in catalogue, not on disk)

## Stale References
[File, line, stale reference, suggested fix]

## Public-Private Cross-Consistency
[Divergences, leaked details]

## User Manual
[Mismatches]

## Recommended Fixes
### Quick (< 2 min each)
### Medium (2-10 min)
### Requires sync script run

Status thresholds:

• OK: 0 issues
• WARN: 1-3 issues
• FAIL: 4+ issues

After the Report

Present the dashboard + top-5 issues. Ask:

• Fix now — apply fixes (count updates, catalogue additions, stale reference removal)
• Run sync — use each repository's documented sync/render command to fix mechanical issues
• Done — report saved

Anti-Patterns

• Do NOT modify any files during the review — report only until explicitly asked to fix
• Do NOT check log/, MEMORY.md, or .context/current-focus.md — these are session artifacts, not documentation
• Do not duplicate a separately requested per-repository prose audit; focus on cross-surface consistency
• Do NOT check code or configuration — this is documentation-only

Cross-References

Skill	Relationship
Installed repository-documentation audit	Per-repo prose, structure, and audience quality.
Repository sync/render command	Mechanical sync of counts, file trees, and generated artifacts.
sync-public-repo	Syncs private → public. Fixes cross-consistency issues.
system-audit	Broader infrastructure health. Sub-agent 4 overlaps on count checks.
update-project-doc	Updates project-level docs (CLAUDE.md, README). Different scope.