Documentation quality assurance

Documentation Quality Assurance

Prevent documentation rot through systematic audits, automated validation, and pipeline-enforced quality gates.

Scope: Inheritable skill. Covers drift detection, preflight validation, semantic accuracy, link integrity, 5-pass quality pipeline, staleness detection, and large-project organization.

Complements: The Documentarian agent uses this skill as its knowledge foundation. This skill is the "what" — the agent is the "who".

Audit Priority: Semantics Over Syntax

The most damaging documentation errors are semantic, not syntactic. A wrong count is annoying. A false claim about functionality is dangerous.

Audit Priority Hierarchy

Priority	Issue Type	Impact	Example
P0 Critical	Phantom features	Users try to use something that doesn't exist	"Entra ID SSO enabled" (code never implemented)
P0 Critical	False security claims	Trust violations, compliance failures	"All data encrypted" (encryption not implemented)
P1 High	Contradictions	User confusion, decision paralysis	README says X, CHANGELOG says Y
P1 High	Stale capability claims	Users miss real features or expect removed ones	Docs describe v3 API but v5 is current
P2 Medium	Broken cross-references	Navigation friction	Link to deleted file
P3 Low	Count drift	Credibility erosion	"109 skills" when there are 123
P3 Low	Formatting issues	Aesthetic concerns	Bad table alignment

Semantic Audit Questions

Before any documentation audit, ask these questions in order:

1. Does this feature actually exist? — Check if documented functionality has corresponding implementation
2. Is this claim still true? — Validate assertions against current codebase state
3. Are there contradictions? — Cross-check related documents for conflicting information
4. Is the version current? — Compare documented versions against package.json/CHANGELOG
5. Do examples work? — Test code snippets against actual API/CLI behavior
6. Do links resolve? — Verify internal and external references (automate this)
7. Are counts accurate? — Check hardcoded numbers against canonical sources (automate this)

Rule: Questions 1-5 require human judgment. Questions 6-7 can be automated. Never spend human attention on automatable checks at the expense of semantic review.

Common Semantic Bugs

Bug Pattern	Detection Method	Fix
Phantom configuration	Grep settings docs, verify in package.json	Remove undeclared settings or add to manifest
Removed feature still documented	Search for deleted code references	Remove or archive documentation
Future feature documented as shipped	Compare roadmap "planned" vs "shipped" markers	Move to correct section
Version mismatch	Regex for version patterns, compare to source of truth	Align all occurrences
Model/API hallucination	Verify external references against official docs	Correct or remove

Count Drift & Docs-as-Architecture (P3)

Count elimination rules, canonical sources, and docs-as-architecture principles are defined in doc-hygiene. Apply those rules during Pass 5 (Lint) of the quality pipeline below.

Document Header Pattern

Comprehensive Metadata Headers

Operational documentation (regression checklists, deployment guides, QA procedures, release workflows) should include comprehensive headers that provide complete context at-a-glance.

Minimal Header (4 lines) — Insufficient:

**Date**: 2026-02-14
**Status**: In Progress
**Purpose**: Verify v5.7.1 UI features
**Method**: Install and test

Enhanced Header (10+ lines) — Comprehensive:

**Version**: 5.7.1
**Date**: 2026-02-14
**Status**: ⚠️ PENDING UI VERIFICATION — WebP avatars regenerated, awaiting restart + testing
**Testing Mode**: CP2 Contingency (Local Install)
**VSIX Size**: 9.44 MB (426 files)
**Key Changes**: Enterprise auth removed, WebP avatars optimized (144×144, 92% reduction)

**Purpose**: Local install verification of all v5.7.1 visual identity + UI features
**Method**: Install VSIX locally, restart VS Code, test in current workspace (CP2 contingency)
**Expected Outcome**: All 9 test sections pass → DoD criterion #4 complete → Ready to publish

Header Field Guidelines

Field	Use For	Example
Version	Software version being documented	5.7.1, v3.2.0-beta
Date	ISO format date of creation/update	2026-02-14
Status	Current state with emoji for scanning	⚠️ PENDING, ✅ COMPLETE, 🚧 IN PROGRESS
Testing Mode	Validation approach or environment	CP2 Contingency (Local Install), F5 Extension Host, Production
Size/Scope	Package size, file count, or metrics	9.44 MB (426 files), 3 breaking changes, 86 tests
Key Changes	What's different in this version	Enterprise auth removed, WebP optimized
Purpose	Why this document exists	One sentence explaining the goal
Method	How the task will be performed	Step-by-step approach or workflow
Expected Outcome	Success criteria	What "done" looks like

Rule: Include enough metadata that anyone can understand the document's context without reading the body. Operational docs reviewed during incidents need at-a-glance clarity.

6-Pass Quality Pipeline

Run these passes in sequence on any document suite:

Pass	Focus	Catches	Type
1. Semantic	Claims match reality, features exist	Phantom features, false claims, contradictions	🧠 Human
2. Architecture	Structural accuracy, diagrams current	Outdated visuals, wrong relationships	🧠 Human
3. Brand	Voice, tone, naming consistency	"Copilot" vs "the AI assistant", passive voice, jargon	🧠 Human
4. Cross-Reference	Link integrity, orphan files	Broken links, unreferenced docs	🤖 Automatable
5. Lint	Formatting, markdown validity, counts	Bad tables, stale numbers, code blocks	🤖 Automatable
6. Consolidation	Redundancy, overlap, merge candidates	Two docs covering same topic	🧠 Human

Rule 1: Don't merge passes — each pass has a single focus. Rule 2: Complete all semantic passes (1-3) before mechanical passes (4-5). A perfectly formatted lie is still a lie. Rule 3: Never allow automated tooling to "pass" a doc suite until human semantic review is complete.

Preflight Validation

Automated Checks

Run before every release or documentation change:

# Example preflight validation script
function Test-DocQuality {
    $errors = @()
    
    # Check 1: All markdown links resolve
    Get-ChildItem -Recurse -Filter "*.md" | ForEach-Object {
        $content = Get-Content $_.FullName -Raw
        $links = [regex]::Matches($content, '\[([^\]]+)\]\(([^)]+)\)')
        foreach ($link in $links) {
            $target = $link.Groups[2].Value
            if ($target -notmatch '^https?://' -and $target -notmatch '^#') {
                $resolved = Join-Path (Split-Path $_.FullName) $target
                if (-not (Test-Path $resolved)) {
                    $errors += "Broken link in $($_.Name): $target"
                }
            }
        }
    }
    
    # Check 2: No orphan files in docs folder
    # Check 3: Required sections present in each doc type
    # Check 4: Version strings match package.json
    
    return $errors
}

Pre-Implementation Cross-Reference Sweep

Before adding any new file, check what already references the concept:

1. Grep for the concept name across all docs
2. Identify which files will need updating
3. Create the new file AND update all references in a single commit

Anti-pattern: Creating a new skill/agent and updating only one reference document. Every catalog, index, and count needs updating simultaneously.

Link Integrity

Link Audit Protocol

Check	How	Frequency
Internal links resolve	Path validation against file system	Every commit
Anchor links work	#heading matches actual heading slugs	Every commit
External links alive	HTTP HEAD request (batch, rate-limited)	Weekly/monthly
Orphan file detection	Files not referenced by any other file	Every release
Circular references	Graph traversal (A→B→C→A)	Quarterly

Orphan Detection

# Find markdown files not referenced by any other markdown file
$allMd = Get-ChildItem -Recurse -Filter "*.md" | Select-Object -ExpandProperty Name
$referenced = @()
Get-ChildItem -Recurse -Filter "*.md" | ForEach-Object {
    $content = Get-Content $_.FullName -Raw
    $links = [regex]::Matches($content, '\]\(([^)]+\.md)')
    foreach ($link in $links) {
        $referenced += Split-Path $link.Groups[1].Value -Leaf
    }
}
$orphans = $allMd | Where-Object { $_ -notin $referenced }

Staleness Detection

Last Validated Dates

Add validation dates to critical documents:

<!-- Last Validated: 2026-02-12 by Documentarian audit -->

Staleness Tiers

Tier	Threshold	Action
Fresh	< 30 days since validation	None
Aging	30-90 days	Flag for review
Stale	90-180 days	Mandatory review before next release
Expired	> 180 days	Consider archival or major rewrite

Drift Detection During Maintenance

During regular maintenance cycles (meditation, dream, release prep):

1. Compare document claims against current reality
2. Flag any numeric drift (counts, versions, dates)
3. Check if referenced files still exist
4. Verify code examples still compile/execute
5. Update Last Validated timestamp after review

Large Project Organization

15+ File Threshold

When a documentation folder exceeds 15 files, introduce numbered chapter folders:

docs/
├── 01-getting-started/
│   ├── installation.md
│   ├── quick-start.md
│   └── configuration.md
├── 02-architecture/
│   ├── overview.md
│   ├── components.md
│   └── data-flow.md
├── 03-operations/
│   ├── deployment.md
│   ├── monitoring.md
│   └── troubleshooting.md
└── README.md              # Index/TOC

Rules:

• Numbered prefixes ensure consistent ordering across all tools
• Each chapter folder has 3-7 files (not 1, not 20)
• Root README.md serves as table of contents with links to all sections
• Flat structure is fine for < 15 files

Multi-Audience Documentation

Audience Matrix

Every doc suite serves multiple readers:

Audience	Needs	Format Preference
New users	Quick start, screenshots, examples	Tutorial (step-by-step)
Experienced users	Reference, API, configuration	Reference (lookup)
Contributors	Architecture, conventions, review process	How-to guides
AI agents	Structured data, clear rules, no ambiguity	JSON > prose, tables > paragraphs

Rule: Each document should declare its audience. A document trying to serve all audiences well serves none of them well.

Ship First, Document After (Threshold)

Document Type	When to Write
User-facing README	Before release
API reference	With the API code
Architecture docs	When design stabilizes
Internal notes	After shipping (retrospective)

Anti-pattern: Blocking a release to write perfect docs. Ship with minimal docs (README + quick start), then iterate.

Doc Audit Checklist

Run this 10-item checklist for any documentation review. Semantic checks first.

Phase 1: Semantic Accuracy (🧠 Human Required)

#	Check	Method
1	Documented features exist	For each feature claim, verify code/config exists
2	No false capability claims	Check "shipped" items against actual implementation
3	No contradictions	Cross-check related docs for conflicting statements
4	Examples work	Copy-paste test critical examples

Phase 2: Mechanical Accuracy (🤖 Automatable)

#	Check	Method
5	All links resolve	Automated link checker
6	No hardcoded counts	Grep for common count patterns
7	Version strings current	Compare against package.json/CHANGELOG
8	File references exist	Verify every referenced file path
9	No orphan files	Cross-reference scan
10	Consistent terminology	Search for variant spellings/names

Rule: Never mark a doc suite "clean" based only on Phase 2 passing. Phase 1 semantic checks are non-negotiable.

TODO Files as Self-Models

A TODO list that contains completed work is worse than no TODO list. TODO.md is a self-model — when read at session start, it forms a mental picture of what exists and what doesn't. Completed tasks masquerading as pending create:

1. Rediscovery tax — work already done gets re-investigated
2. False urgency — energy directed at "building" something already built

The Fix: Done Section First

## Done — Audited [date]
- [x] secretScanner.ts ported to shared/utils/
- [x] All 15 extension.ts files implemented

## Next
- [ ] npm run compile — verify TypeScript
- [ ] F5 smoke test in Extension Development Host

Maintenance Rule: During every meditation or sprint transition, audit TODO.md first. Move completed items to Done. A stale self-model wastes more time than the audit costs.

CHANGELOG Best Practices

Practice	Why
One entry per user-visible change	Users scan, not read
Link to relevant docs/issues	Traceability
Group by: Added, Changed, Fixed, Removed	Consistent scan pattern
Version header matches package.json	No version drift
Date in ISO format (YYYY-MM-DD)	Unambiguous globally
Most recent version at top	Users want latest first