Document Package: make an internal package self-explaining to any model

Most internal packages ship with no usage docs, so every consumer: a teammate, or an AI: reads the source, guesses the intended usage, and gets the gotchas wrong. This skill fixes that: one colocated usage doc per public unit, capturing the API and the non-obvious rules, so the next reader is correct on the first try. It applies to any package: a UI component kit, a utils library, a services/API layer, a hooks package, an internal SDK.

Why it matters for portability: these docs are the layer that makes your package understandable to any model. If you move from one AI tool to another, the new model still understands your package immediately: the knowledge lives in the repo, not in one model's head.

Ask first: always

This writes files into the user's repo. Confirm before doing anything: "I can generate AI-friendly usage docs for <package> so any model (and teammate) understands it correctly: one doc per public unit, with the gotchas. Want me to? (I'll show one sample first.)" Show a sample doc for one unit and get a thumbs-up before fanning out across the package.

Method

1. Discover the public units. Read the package's public entry (index.ts/exports) for the real list, components, exported functions, hooks, services, classes. Note any existing docs' style and match it.

2. Never write a line you have not read the source for. For each unit: the implementation, its types, its variants/options, its tests/stories, and one real usage in the codebase. Verify every claim against the code (never document a guess; mark "unverified" or omit).

3. Write a colocated doc (a README.md/doc beside the unit) with a consistent template:

   • Title + one-line purpose + what it's built on.
   • Signature / API: a table: params/props/args · type · default · description; what it returns.
   • Quick start: the minimal correct usage (real import path).
   • Variants / options / states (where applicable), with a code example each.
   • Composition: how it combines with siblings.
   • Examples: the handful of real scenarios people actually need.
   • Gotchas: the highest-signal section: rules the API doesn't enforce but people get wrong (default values, which state hides what, reserved-but-unimplemented options, ordering constraints).
   • Errors / accessibility / TypeScript: as relevant to the unit's kind.

4. Prioritize the gotchas. The API table can be inferred from types; the gotchas cannot. That's the value.

5. Stamp it so drift is detectable. The code is the SSOT; the doc is derived: so record what it was derived from, the way engineering/ROUTER.md records a hash per node. End each doc with a footer naming the source file(s), the repo commit SHA at generation time, and a short content hash of each source (git rev-parse --short HEAD, git hash-object <file>):

   <!-- generated-from: src/Button.tsx@a1b2c3d (hash 9f4e21bc) · regenerate if the hash differs -->
   

   Detecting drift is then a one-liner anyone (or any model) can run: re-hash the source and compare to the footer, git hash-object src/Button.tsx, or git log a1b2c3d..HEAD -- src/Button.tsx to see whether the unit changed since. Mismatch → treat the doc as stale: re-read the source and regenerate that unit before trusting it. No build script required; if the repo already has a docs check or pre-commit hook, wire the same comparison into it rather than inventing a second mechanism.

Guardrails

• Derive from source, never invent. Unconfirmable behavior → "unverified" or omit.
• One doc per unit, colocated: found next to the thing it describes, travels with it.
• Match the package's existing doc style if any exists; consistency beats your template.
• Confirm scope + a sample before mass-generating: the template must fit before you fan out.