Docs say one thing, code does another. Documentation Audit is a Claude Code plugin that extracts every verifiable claim from your markdown files and checks each one against the codebase. It catches dead file references, wrong defaults, missing env vars, and outdated CLI flags, then expands on what it finds to surface patterns you’d miss by hand.
Install
/plugin install documentation-audit@2389-research
Then invoke with “audit docs”, “verify documentation”, or “check if docs are accurate”.
What it does
The plugin runs a two-pass audit on your project’s user-facing documentation.
Pass 1: Direct Extraction. Each markdown file gets scanned for claims — file paths, config defaults, environment variables, CLI commands, behavioral statements. Every claim is verified against the codebase and given a verdict: TRUE, FALSE, or NEEDS_REVIEW. Extraction runs in parallel (one agent per document) so large doc sets don’t bottleneck.
Pass 2A: Pattern Expansion. False claims from Pass 1 get grouped by type. The plugin then searches all docs for similar patterns. If it finds one dead script reference, it looks for all script references. If it finds one wrong timer interval, it checks every interval claim. This second pass catches 10-20% more issues that a single scan misses.
Pass 2B: Gap Detection. Compares what’s actually in the codebase against what’s documented. Finds scripts that exist but aren’t mentioned, and docs that reference files that were deleted.
Claim types and verification tiers:
- Auto-verified: file references, config defaults, env vars, CLI flags
- Semi-auto: symbol references, version requirements
- Human review: behavioral claims (“runs every 2 minutes”), constraints
How it works
The plugin runs inside an isolated Plan agent context, keeping extraction artifacts out of your main conversation. Each document gets its own parallel Task agent for extraction. Results are aggregated into a single audit report at docs/audits/AUDIT_REPORT_YYYY-MM-DD.md.
The report includes an executive summary with accuracy rates, a table of false claims with line numbers and suggested fixes, a pattern analysis showing root causes (e.g., “9 dead scripts all from the same refactor”), and a human review queue for claims that can’t be auto-verified.
Requirements
- Claude Code with plugin support
- A project with markdown documentation to audit