Evidence-backed docs for humans and agents, written to the Google developer documentation style guide. Classify first. Gate writes. Prefer durable cross-refs over code dumps.
Flow
UNDERSTAND → RESEARCH → CLASSIFY → OUTLINE GATE → WRITE → STYLE → VERIFY
UNDERSTAND names the deliverable, audience, and target paths. Compress when the request already names targets and type. Expand when claims need verification. A copyedit request starts at STYLE. Answer a single-term question ("is allows you to okay?") straight from assets/google-word-list.tsv — quote the guidance and stop.
Rules
- Verify claims in the repository before asserting them. Invented commands, paths, APIs, and env vars are the one unrecoverable failure — omit or mark "Not verified in repository" instead.
- Pick one mode and load its routes before writing.
- Gate creates and overwrites unless the requester approved the targets this turn; a copyedit of a named file carries its own approval. Touch only the files they named — propose the rest.
- Apply the style defaults in
references/style-index.mdto every line you write or edit, and name the rule when you change someone else's wording. - The style references are a snapshot of Google's guide, and every one links the pages it restates. IF someone disputes a rule, the pack doesn't carry it, or the wording carries risk (trademark, product name, legal claim, security claim) → THEN open the live page with a web tool, quote it with its URL, and fix the reference when it disagrees; say so when no fetch was possible.
- A style pass changes wording, not claims; a fact change goes back to RESEARCH.
AGENTS.mdis an index of links and non-obvious rules, not a content dump.- Prefer durable pointers (module path, contract name, doc link) over line numbers and pasted code.
- One Diátaxis type per page; link siblings instead of mixing.
- IF the project documents its own style guide, or the repository already applies a convention consistently → THEN follow it and report the conflict instead of adding a second scheme.
Stop when: outline gate awaits answer; write+style+verify finishes; a word-list lookup answered the question; a missing fact makes the doc dishonest to write (otherwise mark "Not verified in repository" and continue); conventions conflict; user cancels.
Workflows — the mode fixes the route order
| Mode | Deliverable | Route order |
|---|---|---|
| agent-docs | AGENTS.md, nested agent instructions, CLAUDE.md symlink |
modes.md → evidence-research.md → agents-md.md → agent-readable.md → write-verify.md → style-lint.mjs |
| human-docs | README, tutorial, how-to, reference, explanation, runbook | modes.md → evidence-research.md → diataxis.md → agent-readable.md → write-verify.md → style-lint.mjs |
| adr | Architecture decision record | modes.md → evidence-research.md → adr.md → write-verify.md → style-lint.mjs |
| codebase-pack | Multi-file docs set | modes.md → plan the file set → gate once → per file: diataxis.md → write-verify.md → style-lint.mjs |
| style-pass | Edited text, or a style review report | style-index.md → the owning style-*.md → style-lint.mjs → style-review.md for a report someone else acts on |
Routes — workflow references
| Phase | When | Read |
|---|---|---|
| CLASSIFY | Choosing mode or audience | references/modes.md |
| RESEARCH | Gathering or verifying repository facts | references/evidence-research.md |
| CLASSIFY | Choosing the Diátaxis type for human-docs, or reviewing one | references/diataxis.md |
| WRITE | Writing or updating agent instruction files | references/agents-md.md |
| WRITE | Recording a decision | references/adr.md |
| WRITE | Cross-refs, density, durability — read before the first line | references/agent-readable.md |
| OUTLINE GATE, WRITE, VERIFY | Outline gate, write steps, style pass, verify checklist | references/write-verify.md |
| STYLE | Any wording, formatting, or terminology question — maps every guide topic to its owner | references/style-index.md |
Routes — style pack, grouped; references/style-index.md owns the per-topic map
| Ask | Read |
|---|---|
| Which reference owns this topic — every guide topic, one row each | references/style-index.md |
| Prose: tone, person, voice, tense, grammar, one specific word, abbreviations, jargon, translation, inclusive terms | references/style-voice.md, references/style-grammar.md, references/style-words.md, references/style-abbreviations.md, references/style-global.md, references/style-inclusive.md |
| Page shape: headings, lists, numbered steps, notices, tables, footnotes, figures, alt text | references/style-structure.md, references/style-procedures.md, references/style-blocks.md, references/style-images.md |
| Mechanics: bold/italic/code choice, capitalization, filenames, markup, punctuation, numbers, dates, units | references/style-format.md, references/style-punctuation.md, references/style-numbers.md |
| Technical text: code font, samples, command syntax, placeholders, example values, UI wording, link text | references/style-code.md, references/style-cli.md, references/style-examples.md, references/style-ui.md, references/style-links.md |
| Claims and reference text: time words, superlatives, product names, trademarks, third-party text, docstrings | references/style-claims.md, references/style-api.md |
| Producing a review someone else acts on, or checking a rule against the live guide | references/style-review.md, references/style-sources.md |
Scripts and assets — what to run, when, and how
| Run | When | How it behaves |
|---|---|---|
scripts/style-lint.mjs <paths> |
At STYLE, before hand-reading prose | Markdown only. ERROR gates, WARN is mechanical, INFO needs judgment. One finding per rule per line; --max-per-rule caps per file (default 20); --only/--skip select rules; --json for machine output. Exit 1 on ERROR, or on WARN with --strict; 2 on bad usage |
scripts/style-lint.mjs --self-test |
After editing any rule | Lints built-in good/bad fixtures so a gate cannot go inert |
scripts/refresh-word-list.mjs --dry-run |
When the word list looks stale | Rebuilds assets/google-word-list.tsv from developers.google.com; it fetches even with --dry-run, and refuses to write a short parse |
assets/google-word-list.tsv |
Answering a single-term question | 597 guide entries with verdict and guidance; quote it and stop |
assets/google-style-pages.tsv |
Finding which reference owns a guide page, or its URL | All 69 pages as slug, title, owner, url; grep -P "^tables\t" answers ownership, and the URL is what you fetch to verify |
Suppression: <!-- style-lint: ignore-file --> skips a file found by recursion, <!-- style-lint: ignore-line rule-id --> mutes named rules on one line. Both are inert inside a code span, so a page can document them. Docstrings, HTML, and UI strings stay hand-checked.
Related
- Pure code or repository evidence with no docs deliverable →
octocode-research; authoring aSKILL.md→octocode-skills. - Full multi-file pack → plan the file set, gate it once, then work file by file.
- Unclear mode → ask once: agent-docs / human-docs / adr / codebase-pack / style-pass. No Octocode → host search tools.
- Measuring whether this skill triggers and holds its routes →
octocode-graph-eval; the runnable sensors here arescripts/style-lint.mjs --self-test(rules still fire) andscripts/refresh-word-list.mjs --dry-run(word-list drift).