Serial Studio — user-facing docs (write, update, review)

The manual competes with the code for the user's trust; a wrong default or a blog-voiced paragraph loses it. This skill covers doc/help/**, README.md, and examples/**/README.md. The AI-facing material (CLAUDE.md, doc/claude/**, the skills) belongs to ss-ai-audit — don't restyle it from here.

Adapted from AgriciDaniel/claude-blog (MIT): its editorial machinery, rescoped from SEO-blog writing to a technical manual. SEO/GEO advice (question headings, statistics-per-word quotas, summary boxes, keyword placement) was deliberately dropped — a manual is not a blog post. E-E-A-T survives the translation: references/eeat-manual.md carries the experience/expertise/authority/trust signals in manual form, trust evaluated first.

Invariants — name the ones that bind before editing

Per the J-space discipline, state 3-5 of these in chat, in your own words, right before the edit:

1. The code is the ground truth. Every default value, range, UI label, menu path, Pro gate, and behavior claim is verified with Grep/Read against core/ / app/src / app/qml before it is written — never copied from another doc or from memory. Procedure and verdict format: references/ground-truth-factcheck.md.
2. Manual voice, not blog voice. Declarative present tense, no tutorial "we/let's", no marketing adjectives, no meta-references ("in this guide"). scripts/documentation-verify.py is the phrase-level contract — run it, don't re-derive it.
3. Linter-clean is necessary, not sufficient. After the linter passes, run the structural pass in references/second-order-tells.md — the tells that survive a vocabulary fix.
4. Structure mirrors the neighbors. Read 1-2 sibling entries from the same help.json section before writing; match their shape (overview first, configuration tables, numbered quick starts, platform notes).
5. New pages register in doc/help/help.json (id, title, section, file) and cross-link in both directions — the page links its relatives, and at least one existing page links back.

Update workflow

1. Read the target entry in full, plus 1-2 siblings from the same help.json section.
2. Extract the claims your edit touches and verify each against code (ground-truth-factcheck.md). A claim you cannot evidence does not go in the doc.
3. Name the binding invariants in chat (3-5, from the list above).
4. Edit, don't rewrite. Targeted edits that keep the entry's existing structure and any passing prose. Keep terminology aligned with the UI strings and Glossary.md.
5. Gate before handoff:
   • python scripts/documentation-verify.py <files> — fix every finding or justify a <!-- doc-verify off --> fence (fences are a review trigger).
   • Second-order structural pass per second-order-tells.md.
6. Counterfactual check: name the claim in your diff most at risk of being wrong and the file:line evidence that it isn't.

Review workflow (read-only)

When asked to review rather than change, report — do not fix unless asked.

1. First-order: run python scripts/documentation-verify.py <scope> and summarize its findings; don't duplicate them by hand.
2. Second-order: structural pass per second-order-tells.md, each hit with file:line and a quoted example.
3. Ground-truth factcheck: extract the checkable claims, verify against code, report VERIFIED / WRONG (+ correct fact) / NOT FOUND verdicts with file:line evidence.
4. Heuristics rubric: score the entry 0-4 on the ten manual heuristics in references/manual-heuristics.md; tag findings P0-P3.
5. E-E-A-T pass: trust first (factcheck verdicts, gating disclosure, scope honesty — any failure is P0), then experience/expertise/authority per references/eeat-manual.md. A page with no worked example, no concrete value, and no edge case is a UI screenshot in prose — P1.
6. Report prioritized P0 → P3. A WRONG factual claim is always P0. Never silently fix.

New-entry checklist

1. Pick the help.json section it belongs to; read that section's entries for shape.
2. Draft with the shape the section uses: # Title (sentence case), ## Overview stating what the feature is and its edition gating (Pro or free), configuration tables, numbered quick starts, platform notes where behavior differs.
3. Every factual claim carries ground-truth evidence you actually read this session.
4. Write the experience signals in, don't bolt them on: at least one worked example with real output, real platform values, the feature's honest limits (eeat-manual.md).
5. Register in help.json; add cross-links both directions; add new terms to Glossary.md if the manual defines them nowhere else.
6. Run both gate tiers (update workflow step 5).

Rules

• Docs-only. If the code looks wrong while fact-checking, say so in chat — never change code from this skill, and never document the intended behavior instead of the actual one.
• Em dashes are allowed sparingly; the linter's density and ---substitute rules are the contract (the upstream blog plugin's "zero em dashes" rule does not apply here).
• sanitize-commit.py runs documentation-verify.py as part of the pre-commit pipeline; passing the gate here means no surprises there.
• Keep the manual's terminology stable: the UI string is the term ("Project Editor", "frame parser", "Setup panel") — no synonym rotation for variety.