Writing habit-hooks docs

Everything in docs/ is user-facing: it onboards a user (install → run → interpret output → configure) or a contributor extending the system. Short, extremely clear, executable examples as the main communication device.

The finished-cut exemplar: git show f79594fd — docs/smell-vocabulary.md went from 211 lines to 13. What survived: the catalogue table, severity semantics, how to propose a smell. What went: per-plugin translation tables, design rationale, edge-case essays.

Properties of a doc

Audience and purpose

• One doc, one reader, one job — user onboarding or contributor extension. A doc serving both does neither well.
• Answers "what do I do", never "how did we decide". Rationale lives in PRs, commit messages and tests; mechanism walk-throughs live with the doc that owns them (architecture.md for the pipeline).
• Contribution sections state acceptance criteria ("name the smell by the problem it creates, not the tool"), not the reasoning behind them.

Examples are the message

• The executable example carries the meaning; prose between examples only supplies concepts and context the example can't show. A paragraph that re-explains what the example demonstrates gets deleted.
• Examples run as written — the spec harness enforces it, which is why examples are the safest device: they can't rot, prose can.
• *.spec.md files are documentation first, tests as a side effect. Trimming prose around cases is a docs change; deleting or editing cases changes test coverage and goes through normal test review, not a docs cut.

Stays true

• Contains nothing code can answer. State inventories — which tool maps to which smell, which plugins ship which sensors — go stale the release after they're written; link to where the truth lives instead of copying it.
• Each fact stated once. Cross-reference instead of restating; duplicated explanations drift apart.

Length

• Every sentence carries a fact the reader acts on. First cut: history, edge-case essays, "deliberate exception" narratives.
• Lead with the organising fact as a definition, in the fewest words. Never advertise what the doc lets you do — that restates the sections and reads as a sales pitch.
• One sitting per doc. A doc that needs a table of contents wants to be two docs — or its headlines to carry their promises, not an index table to carry them.

Format

• No hard newlines inside a paragraph; let the editor soft-wrap.

Boundaries

• Core docs are plugin-agnostic, same rule as the core code (see AGENTS.md). Per-plugin detail lives in the plugin.
• Concepts defined before first use. "Sensor", "mapper", "smell" mean the same thing in every doc.

The process — writing or fixing any doc

Run these in order, before writing prose or cutting it. A doc that skips a step is why a reader bounces off it (docs/config.md before its 2026-09 rework: no audience, no promise, mechanism-named sections).

1. Decide who the audience is. One doc, one reader, one job. Signal it by talking to that reader — never announce who the doc is for, and never narrate what the reader just did ("you open this page to…").
2. Phrase the reader's expectation on opening it — one sentence: "I will find out how to disable a rule I don't agree with." If it takes two sentences, the doc has two jobs: split it.
3. Make the headlines deliver on those expectations. Name the problem being solved, never the mechanism — not the design name (## Advanced configuration) and not the syntax ([smells.<name>]): ## Silence or demote a smell. If a headline can't carry its promise, rename the headline — don't paper over it with a task index.
4. Make the first paragraph summarise the content and place the doc in context. The summary itself tells the right reader they are in the right place.
5. Write each section to deliver on the promise of its headline — and only that promise. Deliver with executable examples; reference data is a bullet list, - **key**: description, never a table and never prose.

Cutting an existing doc

Fix structure first with the process (audience → expectation → headlines → first paragraph → sections), then cut prose.

1. Delete every sentence whose absence changes nothing the reader does or understands.
2. Delete state inventories code already answers; link to the truth instead.
3. Move per-plugin detail to the plugin, and mechanism walk-throughs to the doc that owns them.
4. If two distinct readers or jobs remain, split into two docs.
5. Keep: reference tables the reader looks up, acceptance criteria, executable examples.

Check: read only the examples — does the doc still teach its job? Then read the prose; anything the examples already said is gone.

Cross-reference other docs with a link, never a summary.

Before handing over

For every line, ask as the reader with a task in hand, not the writer who wrote it: does the user need this — and is this the one place it is said? Repeated anywhere else, or nothing the user does depends on it → delete.

Intuitive rules (precedence, who-wins) are stated once, at the end, for the reader checking their intuition — never inline at each point of possible conflict.

Examples that demonstrate a default show nothing; the default is already stated. Delete them.