Skill: Validate Documentation

What This Skill Does

Answers the question: "Are my docs still accurate?" — without reading a single source file.

Uses git metadata (timestamps, diff stats, file lists) to detect which documentation is stale and which is current. Produces a structured staleness report that tells update-docs exactly which modules to target.

This is the CHECK path between generate-docs (create) and update-docs (update). Without it, update-docs must perform an expensive full-scan of all docs and all source files to determine what changed.

When to Use

Before running update-docs — to scope the update to only stale modules
At the start of a session — to understand documentation health before working
When the user asks "are my docs up to date?" or "what's stale?"
As part of smart-start — automatically invoked during session bootstrap
After a series of commits — to check if documentation needs attention

Do NOT use this skill to actually update documentation — use update-docs with the staleness report as input.

Execution Model

Always: the primary agent runs this skill directly.
Rationale: validate-docs is cheap (git commands + frontmatter reads, ~2-3k tokens total). Running it in a subagent would add delegation overhead that exceeds the skill's own cost.
No subagent needed: this skill never reads source files and never writes to docs/.
Output: chat-based report (not written to a file). The report is ephemeral — it reflects a point-in-time assessment.

Why This Skill Exists

Without validate-docs, there are only two options:

Run update-docs speculatively — reads ALL docs + ALL source files to find what changed. Cost: 20-50k tokens for a medium project. Often 80% wasted because only 2 of 8 modules were stale.
Skip updates — docs drift out of sync. Future sessions work with incorrect context, leading to wrong implementation decisions.

validate-docs provides the missing middle ground: a ~2-3k token check that produces a targeted action list.

Workflow

Step 1: Discover Doc Inventory

List all documentation files under docs/:

List docs/modules/ for module documentation files
List docs/features/ for feature documentation files
Check for docs/overview.md

If docs/ does not exist or is empty, report "No documentation found" and suggest generate-docs. Stop here.

Step 2: Determine Doc Timestamps

For each documentation file, get its last modification date from git:

git log -1 --format=%aI -- docs/modules/<name>.md

This gives the ISO timestamp of the last commit that touched the doc file. This is more reliable than frontmatter version fields, which may not be updated consistently.

If a doc file has never been committed (new, untracked), treat it as "just created" (current timestamp).

Step 3: Extract Source Mappings

For each module documentation file, read only the ## Structure section (not the entire file). Extract the source paths listed in the Structure table.

## Structure

| Path               | Type | Purpose        |
|--------------------|------|----------------|
| src/auth/          | dir  | Auth module    |
| src/auth/handler.ts| file | HTTP handlers  |

From this, derive the module's source scope: src/auth/ in this example.

Token optimization: read only the lines between ## Structure and the next ## heading. Do not read the entire module doc. This typically costs ~50-100 tokens per module instead of ~300-600 for the full file.

If the Structure section is missing or unparseable, fall back to checking the module name against common directory patterns (src/<module>/, lib/<module>/, packages/<module>/, <module>/).

Step 4: Check Source Changes

For each module's source scope, query git for changes since the doc was last updated:

git log --since="<doc_timestamp>" --oneline --stat -- <source_path>

Parse the output to determine:

Number of commits since doc update
Files changed with insertion/deletion counts
Authors involved (optional, for context)

If the output is empty → module docs are current. If the output has commits → module docs are stale. Capture the commit subjects and changed files for the report.

This is the core optimization: we never read source files. Git already knows what changed.

Step 5: Check Overview Staleness

Check if the project overview needs updating:

New modules: Compare the module list in docs/overview.md (from the ## Modules table) against actual directories in the project. Any directory that looks like a module but isn't documented → overview is stale.
Removed modules: Any module referenced in the overview that no longer exists in the filesystem → overview is stale.

Structural changes: Check for significant project-level changes since the overview was last updated:

git log --since="<overview_timestamp>" --oneline -- package.json pyproject.toml go.mod Cargo.toml pom.xml Makefile docker-compose*.yml

Step 6: Assess Feature Doc Staleness

Feature docs reference modules. Use the module staleness results from Step 4 to infer feature staleness:

Read only the ## Implementation section of each feature doc (the table that maps modules to symbols).
If any referenced module is stale → mark the feature as "Review Recommended" (not definitively stale, since the module change may not affect the feature).

This avoids any source file reads for feature validation.

Step 7: Present the Report

Present the staleness report in chat using the format defined below. Do NOT write the report to a file — it is ephemeral session context.

If smart-start invoked this skill, return the report data for integration into the smart-start assessment.

Report Format

The report follows this structure:

## Documentation Validation Report

Generated: <timestamp>

### Summary

| Status | Count |
|--------|-------|
| Current | N |
| Stale   | N |
| Missing  | N |

### Module Documentation

| Module | Status | Commits Since Update | Key Changes |
|--------|--------|----------------------|-------------|
| <name> | Current | – | – |
| <name> | Stale | N commits, M files | `file1` (X+/Y-), `file2` (X+/Y-) |
| <name> | Missing | new module | `<path>` (N files) |

### Feature Documentation

| Feature | Status | Reason |
|---------|--------|--------|
| <name> | Current | all referenced modules current |
| <name> | Review Recommended | depends on stale module: <module> |

### Overview

| Check | Status | Detail |
|-------|--------|--------|
| Module list | Current/Review Recommended | <detail> |
| Tech stack | Current/Review Recommended | <detail> |

### Recommended Actions

1. <prioritized action, e.g. "Run update-docs targeting modules: auth, notifications">
2. <next action>

Integration with Other Skills

Skill	Relationship
`update-docs`	validate-docs BEFORE update-docs. Pass the staleness report as input so update-docs targets only stale modules.
`generate-docs`	If validate-docs finds no `docs/` directory, suggest generate-docs instead.
`smart-start`	smart-start invokes validate-docs as part of its session assessment.
`resume-plan`	validate-docs can run alongside resume-plan to add documentation health to the session briefing.

Rules

Never read source files: This skill uses git metadata only. If you find yourself reading .ts, .py, .go or similar source files, you are doing it wrong. The entire point is to avoid source file reads.
Minimal doc reads: Read only frontmatter and specific sections (Structure, Implementation tables). Never read full documentation files.
Git is the source of truth: Use git log, git diff --stat, and git ls-files for change detection. Do not rely on filesystem timestamps or frontmatter version fields.
Report is ephemeral: Do NOT write the report to a file. It is presented in chat and used by the primary agent to decide next actions.
No false positives over false negatives: It is better to flag a module as "stale" when it might be current (a commit touched the source path but didn't change documented behavior) than to miss a genuinely stale module.
Suggest, don't execute: This skill produces a report and recommendations. It does NOT update any documentation files. That is update-docs's job.
Fast execution: The entire skill should complete in under 10 tool calls. If you're making more calls than that, you're over-engineering the check.
No built-in explore agent: Do NOT use the built-in explore subagent type.