Audit — Documentation Quality & Consistency Review

Systematically review existing project documentation for quality, completeness, consistency with code, and internal coherence. Produce a prioritized findings report.

Core Principles

1. Evidence-based: Every finding must cite a specific file, section, or line — no vague complaints
2. Code-grounded: Cross-reference documentation claims against the actual codebase
3. Prioritized: Findings are classified by severity so the user can fix what matters first
4. Read-only by default: This skill only reads and reports — document modifications require explicit user opt-in (Step 8)
5. Honest: Report real issues, don't inflate findings to look thorough

Severity Levels

Workflow

Step 1: Determine Audit Scope

Parse the arguments to determine what to audit:

1. If a path argument is provided (e.g., /spec-forge:audit ../../other-project), use that as the project root
2. If no path, use the current working directory
3. Verify the target has documentation (check for docs/, README.md, or markdown files)

If no documentation found, inform the user and stop.

Use AskUserQuestion to ask:

• Audit focus: Full audit, or focus on a specific area? (Options: Full / API docs only / Feature docs only / Architecture docs only / Custom path)
• Code alignment: Should I cross-reference docs against the codebase? (Yes / No — skip if docs-only repo like aipartnerup-docs)

Step 2: Document Inventory

Scan the documentation landscape and build a complete inventory.

Fast path (preferred): resolve <sf_scripts> (see @../shared/scripts.md) and run

python3 "<sf_scripts>/sf-scan.py" --root "<project_root>" --docs-only

Use its documents / document_index fields as the starting enumeration — the discovered markdown files plus the requirement IDs each declares — so you do not have to glob and count by hand. The script enumerates; you still judge each document's TYPE (the categories below) and its quality. Fallback: if python3 is unavailable or the script is not found, glob and enumerate by hand as described below.

1. Glob for all markdown files in the docs directory (and README at root)
2. Categorize each document by type:
   • api — API reference documentation
   • architecture — Architecture, design, system overview
   • feature — Feature specifications or descriptions
   • guide — How-to guides, tutorials, getting started
   • spec — Protocol specs, conformance, standards
   • decision — ADRs, decision records
   • reference — Data models, type mappings, glossaries
   • meta — READMEs, indexes, navigation docs
   • other — Doesn't fit above categories
3. Read each document — read enough of each to extract its metadata and claims; for very large files read strategically (headings + relevant sections) rather than a fixed prefix. (The sf-scan enumeration above already supplies the file list and declared IDs, so a fixed first-200-lines metadata pass is unnecessary.)
4. Record metadata for each document:
   • File path
   • Category
   • Approximate word count
   • Last modified date (from git if available)
   • Key topics covered
   • Cross-references to other docs (internal links)
   • Cross-references to code (file paths, class/function names mentioned)

Display the inventory:

Document Inventory: {project-name}
  Path: {project-root}
  Total documents: {N}

  Category          | Count | Files
  api               |   4   | api/README.md, api/executor-api.md, ...
  feature           |   8   | features/acl-system.md, ...
  guide             |   7   | guides/creating-modules.md, ...
  ...

Step 3: Code Alignment Check

Skip this step if the user opted out in Step 1 (e.g., docs-only repo).

Cross-reference documentation against the actual codebase:

3.1 API Surface Audit

1. Scan the codebase for public API surfaces: exported functions, classes, interfaces, types, endpoints
2. Compare against API docs: For each documented API, verify:
   • Does the function/class/endpoint still exist in code?
   • Do parameter names and types match?
   • Do return types match?
   • Are documented examples still valid?
3. Find undocumented APIs: Public APIs that exist in code but have no documentation

Report findings:

• Stale API docs: Documented APIs that no longer exist or have changed
• Undocumented APIs: Code APIs with no corresponding documentation
• Parameter mismatches: Documented params that don't match code signatures

3.2 Feature Coverage Audit

1. Identify features in code: Major modules, systems, or functional areas in the codebase
2. Compare against feature docs: Is each significant feature documented?
3. Check feature doc accuracy: Do feature descriptions match the current implementation?

Report findings:

• Undocumented features: Significant code modules with no feature documentation
• Stale feature docs: Feature docs describing behavior that has changed
• Missing feature aspects: Feature docs that exist but skip important aspects (config, error handling, limitations)

3.3 Architecture Alignment

1. Infer architecture from code: Module structure, dependency graph, communication patterns
2. Compare against architecture docs: Does the documented architecture match reality?
3. Check for architectural drift: Places where code has evolved away from the documented design

Step 4: Internal Consistency Check

Check documents against each other for contradictions and coherence:

4.1 Cross-Document Consistency

1. Terminology consistency: Same concept referred to by different names across docs?
2. Factual consistency: Do different docs make contradictory claims? (e.g., architecture doc says "microservices" but feature doc says "monolith")
3. Version/dependency consistency: Do different docs reference different versions of the same thing?
4. Link integrity: Do internal cross-references actually point to existing documents/sections?

4.2 Structural Consistency

1. Format consistency: Do similar documents follow similar structures? (e.g., all feature docs should have similar sections)
2. Depth consistency: Are some areas documented exhaustively while others are skeletal?
3. Naming consistency: File naming conventions, heading styles, code formatting

Step 5: Quality Assessment

Fast path (preferred) for the structural half of Completeness: resolve <sf_scripts> (see @../shared/scripts.md) and run, per document,

python3 "<sf_scripts>/sf-verify-doc.py" "<file>" --type unknown

Audit runs on arbitrary external doc collections, not spec-forge's own doc types, so pass --type unknown to run only the type-agnostic checks — non-empty + titled, well-formed/unique IDs, and no leftover template placeholders — instead of expecting spec-forge sections. Use its findings for the deterministic part of the Completeness dimension below (TBD/TODO markers, placeholder text, empty sections). Whether a doc is missing sections for its own purpose stays a model judgement. Fallback: if python3 is unavailable or the script is not found, check these by hand.

Evaluate each document (or document category) on quality dimensions:

Step 6: Generate Findings Report

Write the findings report to {project-docs-path}/audit-report.md (or a user-specified path).

Report format:

# Documentation Audit Report

> Project: {project-name}
> Audited: {date}
> Scope: {full / api-only / feature-only / etc.}
> Documents reviewed: {N}
> Code alignment: {Yes / Skipped}

## Executive Summary

{2-3 paragraph overview: overall documentation health, biggest strengths, most critical gaps}

## Findings Summary

| Severity | Count | Categories |
|----------|-------|------------|
| Critical |   {n} | {which areas} |
| Major    |   {n} | {which areas} |
| Minor    |   {n} | {which areas} |
| Info     |   {n} | {which areas} |

## Document Health Matrix (optional)

Grade a document only where the letter grade drives a priority action; otherwise summarize its health in prose. The A–D boundaries are undefined, so do not force a grade onto every document just to fill the table — a clear sentence beats pseudo-precision. Include only the rows worth grading.

| Document | Completeness | Accuracy | Clarity | Currency | Overall |
|----------|-------------|----------|---------|----------|---------|
| {path}   | {A/B/C/D}   | {A/B/C/D}| {A/B/C/D}| {A/B/C/D}| {A/B/C/D} |

(A = Excellent, B = Good, C = Needs Work, D = Poor)

## Critical Findings

### CRIT-001: {Finding title}
- **Location**: {file path, section/line}
- **Issue**: {What is wrong}
- **Evidence**: {How you know — code reference, conflicting doc, etc.}
- **Impact**: {Why this matters}
- **Fix**: {Concrete recommendation}

### CRIT-002: ...

## Major Findings

### MAJ-001: {Finding title}
...

## Minor Findings

### MIN-001: {Finding title}
...

## Observations & Suggestions

### INFO-001: {Suggestion title}
...

## Coverage Map

{Mermaid diagram showing documentation coverage — which areas are well-documented vs. gaps}

## Recommended Priority Actions

1. **{Action}** — fixes CRIT-001, CRIT-002 — {estimated effort: small/medium/large}
2. **{Action}** — fixes MAJ-001 through MAJ-003 — {effort}
3. ...

Step 7: Present Results

Display a summary to the user:

Audit complete: {project-name}

  Documents reviewed: {N}
  Findings: {critical} critical, {major} major, {minor} minor, {info} info

  Report: {path-to-audit-report.md}

  Top 3 priority actions:
    1. {action} — {severity} — {effort}
    2. {action} — {severity} — {effort}
    3. {action} — {severity} — {effort}

Use AskUserQuestion to ask:

• Fix mode: Would you like me to fix any of these issues now?
  • Fix critical findings
  • Fix all findings (critical + major + minor)
  • Just the report, I'll fix manually
  • Fix specific findings (list IDs)

If the user wants fixes, proceed to Step 8.

Step 8: Apply Fixes (Optional)

For each finding the user wants fixed:

1. Read the target document
2. Apply the fix described in the finding
3. Mark the finding as resolved in the report

Rules for fixes:

• Only modify the specific section cited in the finding
• Do NOT restructure or rewrite entire documents
• Do NOT add new documents — only fix existing ones
• If a fix requires information you don't have (e.g., the correct API signature), mark it as "needs manual review" instead of guessing
• After all fixes, re-run a quick validation to ensure fixes didn't introduce new issues

Update the audit report to mark resolved findings:

### CRIT-001: {Finding title} [RESOLVED]

Notes

1. Scope flexibility: The audit can target the current project, a sibling project, or any directory with docs
2. Incremental audits: If a previous audit-report.md exists, note which findings are new vs. recurring
3. No false positives: It's better to miss an issue than to report a non-issue. Every finding must have concrete evidence
4. Respect existing structure: The audit evaluates documentation as-is — it doesn't impose spec-forge's own document structure on projects that use different conventions