document

Produce documentation a reader can actually use, partitioned by what they need — not one undifferentiated wall of prose.

1. Scope & intent. Confirm what to document and for whom. Ask once whether output should be inline (docstrings/README sections) or standalone files under docs/ — default to standalone. Read ./CONTEXT.md if present for the stack, domain language, and where docs already live.
2. Read the code first (non-negotiable). Before writing a line, read the implementation, its tests, config, and the public surface. Documentation is derived from what the code does, not what it should do. If something is undocumented because it's unclear, say so — don't invent it.
3. Partition by Diataxis. Decide which of the four quadrants the subject actually needs (not everything needs all four):
   • Reference — complete, accurate, factual description of the surface (API, flags, config). Information-oriented. Tie every entry to the code.
   • Explanation — why it works this way: design decisions, trade-offs, alternatives rejected. Understanding-oriented. (If a decision is load-bearing, suggest capturing it via the adr skill.)
   • How-to — task-oriented recipes that assume basic familiarity ("how to add a new handler").
   • Tutorial — a learning-oriented walkthrough for a newcomer, with a working example that produces a visible result early. Every step must actually run.
4. Write reference first, then explanation, then how-to, then tutorial. Reference grounds the rest; the others link back to it.
5. Cross-link and index. Connect the quadrants to each other and update the README / docs index so the new pages are discoverable. Don't leave orphans.
6. Self-review against the code. Check every factual claim against the implementation, confirm examples run (macOS-first commands), and keep the voice plain — apply the humanizer patterns to your own output. Markdown by default.

Don't pad, don't duplicate the same content across quadrants, and don't auto-commit — present the files for review.