review-documentation

Enter planning mode. Audit the three documentation pages for mechanical errors and content staleness. Use your judgment on what matters — the goal is a prioritized report the human can act on, not an exhaustive lint.

Scope

The three living documentation pages:

• README.md — product page (features, install, config, usage)
• docs/what-autohotkey-can-do.md — AHK technical showcase
• docs/llm-development.md — Claude Code workflow and guardrail system

Excluded: docs/making-my-unteachable-ai-build-its-own-cage.md is a published article and time capsule — do not audit it.

Phase 1 — Mechanical Verification (High Confidence)

These are objectively right or wrong. No judgment needed.

Link integrity

Every markdown link ([text](path)) in all three pages. Verify each target exists:

• Relative file links (../src/gui/d2d_shader.ahk, ../tools/mcode/build_mcode.ps1, etc.)
• Cross-doc links (what-autohotkey-can-do.md, llm-development.md, etc.)
• External URLs — just verify they're well-formed, don't fetch them
• Anchor links if any

Numbers and counts

Verify claims against the actual codebase. Use the most efficient method for each:

Don't obsess over exact matches — "200+ settings" is fine if the count is 215. Flag when claims are meaningfully wrong (says 183 shaders but there are now 195, says 71 checks but there are now 78).

Phase 2 — Feature Coverage (Medium Confidence)

This requires judgment. The goal is to catch features that exist in code but aren't mentioned in documentation, or features mentioned in docs that no longer exist.

README feature gaps

Scan src/shared/config_registry.ahk for top-level config sections. Each section roughly maps to a feature area. Compare against what the README mentions. New sections since the README was last written likely represent missing features.

Also check:

• New shader categories or significant shader count growth
• New command-line flags in alt_tabby.ahk
• New diagnostic log types
• New keyboard shortcuts or user-facing behaviors

README stale features

Check if any features described in the README reference config keys or behaviors that no longer exist. Use query_config.ps1 -Usage <key> for specific keys mentioned.

AHK page — new showcase opportunities

Look for technically impressive additions not yet covered. Potential signals:

• New files in src/gui/, src/core/, src/shared/ that represent significant new subsystems
• New Win32 APIs being called via DllCall (grep for DllCall in recently modified files)
• New MCode modules beyond icon_alpha.ahk
• Growth in the shader pipeline (new shader types, new cbuffer fields, new compute patterns)
• New build pipeline stages in compile.ps1

Don't just flag everything new — use judgment on whether it's "wait, AHK does that?" material or routine development.

LLM page — new guardrail highlights

Look for growth in the tooling that the LLM page should reflect:

• New check scripts or sub-checks not accounted for
• New query tools
• New skills worth calling out (especially review skills with novel methodologies)
• Changes to the pre-gate architecture
• New MCP integrations or context management patterns

Phase 3 — Cross-Page Consistency (Low Priority)

Check that claims made in multiple pages agree:

• "By the Numbers" tables in both the AHK and LLM pages should have consistent values
• The README's "Behind the Scenes" summary should match what the linked pages actually describe
• Feature descriptions in README should be compatible with technical descriptions in the AHK page

Explore Strategy

Run in parallel by page — each page is independent:

• Agent 1: README mechanical checks (links, feature coverage scan)
• Agent 2: AHK page mechanical checks (links, numbers, new showcase candidates)
• Agent 3: LLM page mechanical checks (links, numbers, new guardrail highlights)

After agents report, do the cross-page consistency check yourself.

Output Format

Organize findings into two categories:

Mechanical Issues

Objectively broken — links, numbers, references.

Content Opportunities

Subjective — features, highlights, or coverage gaps worth considering. Include enough context for the human to make a quick yes/no decision.

Order by impact: broken links first, stale numbers second, missing features third, nice-to-haves last.

Ignore any existing plans — create a fresh one.