Dedao Write

dedao-write — Eight-Role Writing Pipeline

Convention: see conventions/quality.md for the MANDATORY citation format, source precedence, and back-link rules.

Convention: see conventions/brain-first.md — check the brain (and the user's own notes) before reaching for external sources.

Convention: see conventions/model-routing.md — each role runs on a cost-appropriate model (see the per-phase Model line below).

Convention: see conventions/subagent-routing.md — when to run a phase inline vs as a Minion (long phases, user walks away).

Filing rule: see _brain-filing-rules.md — finished articles file under writing/.

What this solves

A good article needs eight different kinds of judgment, and one person (or one undifferentiated LLM pass) does all of them badly at once: picking the right angle, gathering the user's own material, finding evidence, drafting, matching the user's voice, checking facts, editing structure, and laying it out. This skill splits the work into eight named roles that run as sequential phases. Each phase has a single job, reads the previous phase's output, and hands a clean artifact to the next.

The defining feature is the voice-polisher (Phase 5): it studies the user's past writing across every store they own — gbrain writing/ + originals/, GetNote saved knowledge, the Obsidian vault, and the cross-model AI archive — so the finished piece reads like the user wrote it, not like an LLM did.

The eight roles (run in order)

Phase 1  Topic Strategist    选题策划   Pick the angle worth writing + lock the spec
Phase 2  Knowledge Steward   知识管家   Pull the user's OWN material AND voice samples
Phase 3  Researcher          调研专家   Add data + stories as evidence for each claim
Phase 4  Writer              写手       Turn scattered material into a complete draft
Phase 5  Voice Polisher      风格打磨师 Rewrite the draft in the user's personal voice
Phase 6  Fact-Checker        事实核查   Verify every fact, number, date, name, quote
Phase 7  Editor-in-Chief     主编       Audit structure, logic, expression, reader value
Phase 8  Layout Designer     排版设计师 Present it with hierarchy for CJK + English

Each phase is a focused pass. Do not collapse them — the value is in the separation. A phase may loop back one step (editor → writer) but never skips.

Model routing (per model-routing.md). Roles differ wildly in cognitive load; do not run all eight on one model:

Inline vs Minion (per subagent-routing.md). Short phases (1, 8) run inline. Phases 3–7 can be long; if the user will walk away or wants progress, run them as Minions. If >3 phases need to run concurrently you would use Minions — but this pipeline is sequential, so concurrency is not a factor here.

Phase 1 — Topic Strategist (选题策划) · Model: Opus

Job: From a fuzzy idea, pick the single most-worth-writing topic, lock a sharp angle, AND fix the article spec (length, depth, channel). Do NOT gather material yet. The spec drives every later role's granularity.

1. Restate the user's raw idea in one sentence.
2. Generate 3–5 candidate angles. For each: who it's for, what it argues, why it's worth the reader's time, what makes it non-obvious.
3. Score each on: does the user have unique material/experience for it (gbrain query for prior takes), is the claim defensible, is the audience real.
4. Recommend ONE. State: working title, one-sentence thesis, intended reader, the promise (what the reader can DO after reading), and the spec: target length (e.g. ~2000 字 / ~1500 words), depth level (overview vs deep), and publish channel (公众号 / blog / newsletter) — channel sets tone + format.

Output: a topic brief — {working_title, thesis, audience, promise, spec:{length, depth, channel}, why_this_angle}.

Human checkpoint (recommended): topic + angle is the highest-leverage human decision. Show the brief and confirm before Phase 2 — unless the user's request already pinned the angle, in which case proceed and note the assumption.

Phase 2 — Knowledge Steward (知识管家) · Model: Sonnet

Job: In ONE retrieval pass, surface BOTH (a) the user's own knowledge on this topic and (b) representative samples of the user's past writing for voice study. Gathering both here means Phase 5 reuses these samples instead of re-scanning the stores a second time.

Query, in this order (brain-first, then the user's other stores):

1. gbrain — the primary store:

   gbrain query "<thesis keywords>" --json            # topic material
   gbrain query "type:writing <topic>" --limit 50 --json
   gbrain query "type:writing" --limit 20 --json       # voice samples (any topic)
   gbrain list --tag <topic> --json
   

2. GetNote — saved knowledge and recordings (MCP): mcp__getnote__recall / mcp__getnote__recall_knowledge on topic keywords; list_topic_notes if a matching topic exists.
3. Obsidian vault — local markdown notes (MCP mcp__obsidian search-vault).
4. AI archive — prior cross-model conversations: ai-search "<topic>" (Claude/Codex/Gemini history, already secret-redacted).

For each store: if it's unavailable or empty, log "store X: no material" and move on — never block the pipeline on one missing store.

Output: a material dossier with two parts — (a) topic material: deduplicated user points, prior phrasings, anecdotes, half-formed takes, each source-tagged (brain slug / GetNote id / Obsidian path / archive file), verbatim where the phrasing is good; (b) voice samples: 3–8 of the user's past pieces (full or substantial excerpts) set aside for Phase 5, so the polisher does not re-query the stores.

Phase 3 — Researcher (调研专家) · Model: DeepSeek/Qwen (cheap, large-context)

Job: For each claim in the thesis, attach supporting evidence — data, studies, concrete stories. The steward gave the user's side; the researcher adds backing.

1. List every assertion the article will make.
2. For each, find a number/data point, a credible source, or a concrete story/example. Prefer the brain first (gbrain query), then external sources only where the brain is thin.
3. Record source metadata for EVERY external fact: {publication, author, url, date} — enough for the fact-checker to re-find it and for Phase 8 to build the citation in the format defined by conventions/quality.md.
4. Flag any claim with no evidence as UNSUPPORTED — the writer softens it or the editor cuts it.

Output: an evidence map — claim → [{type, content, source:{publication, author, url, date}}], with UNSUPPORTED flags where backing is missing.

Phase 4 — Writer (写手) · Model: Opus

Job: Turn the scattered material (dossier + evidence map) into one complete, coherent draft sized to the Phase 1 spec. This is where zero-in-on-structure happens — pull the pieces into a single continuous expression.

1. Build a working outline from the thesis sized to spec.length and spec.depth: opening hook → argument sections → close. Each section maps to claims + their evidence.
2. Title + opening hook get a dedicated pass. Draft 2–3 title options and a first paragraph that earns the reader's attention; these carry the most weight and are worth isolating from body drafting.
3. Write the full draft in continuous prose. Weave the steward's user-material and the researcher's evidence into each section. Reuse the user's own phrasings from the dossier verbatim wherever they fit (not paraphrased).
4. Every external fact carries an inline citation marker tied to the evidence map (final citation formatting is applied in Phase 8 per quality.md).
5. Do NOT polish voice and do NOT fact-check yet — later phases. Goal here is completeness, correct length, and logical flow, not finish.

Output: a complete first draft (markdown) sized to spec, with title options, a worked opening, section structure, and inline citation markers. Coverage check: every non-UNSUPPORTED claim appears; every piece of evidence is used or consciously dropped.

Phase 5 — Voice Polisher (风格打磨师) · Model: Opus

Job: Rewrite the draft so it reads in the user's personal voice. The signature role. Learn the style from the voice samples Phase 2 already gathered (do not re-query the stores), build a fingerprint, then apply it.

1. Learn the voice from the Phase 2 voice samples. Extract a concrete fingerprint: sentence length and rhythm, paragraph shape, favored connectors and transitions, punctuation habits (em-dash vs comma, CJK full-width usage), how the user opens and closes, level of directness, recurring metaphors, vocabulary tells, and what the user never does.
2. Apply it. Rewrite the draft to match the fingerprint. Preserve every fact, number, citation marker, and structural beat — change only voice, rhythm, word choice, and transitions. Prefer the verbatim user-phrasings from the dossier.
3. If Phase 2 surfaced fewer than ~3 voice samples, say so explicitly, apply a light touch, and ask the user for 1–2 reference pieces rather than inventing a voice.

Output: the draft rewritten in the user's voice, plus a short style_fingerprint note recording what was matched (so later runs are consistent). Facts and citation markers untouched.

Phase 6 — Fact-Checker (事实核查) · Model: Sonnet

Job: Verify every fact, number, date, name, and quote. Trust nothing that can't survive scrutiny.

1. Enumerate every checkable claim: numbers, dates, names, quotes, attributions, "studies show" statements.
2. Re-verify each against its source in the evidence map. Re-fetch external sources where needed (use the web/fetch tooling available in the session, e.g. firecrawl MCP or ai-search for prior captures — this skill's declared gbrain tools cover brain re-verification; web re-fetch uses session tooling). A quote must match verbatim; a number must match the cited figure; a date and a name must be exact.
3. Mark each: VERIFIED (with source), CORRECTED (old → new + source), or UNVERIFIABLE (no source survives — flag for the editor to cut or soften).
4. Confirm every external claim carries a citation marker; add missing ones or flag the claim.

Output: a fact-check report — every checkable claim with its verdict and source — plus a corrected draft with every CORRECTED fix applied inline. Never let an UNVERIFIABLE claim pass silently toward publication.

Phase 7 — Editor-in-Chief (主编) · Model: Opus

Job: The gate before publication. Audit the whole piece — structure, logic, expression, AND whether it delivers the Phase 1 promise to the reader.

1. Structure: does the opening earn attention, do sections flow, does the close land?
2. Logic: does each argument follow, are there gaps or non-sequiturs, does any UNSUPPORTED/UNVERIFIABLE claim still need cutting or softening?
3. Reader value: measure the draft against the Phase 1 promise — can the intended reader actually DO what the article promised after reading? If not, name the gap. This is the acceptance test, not a vibe check.
4. Expression: cut redundancy, tighten flabby sentences, fix anything fighting the user's voice — without re-introducing facts the checker removed.
5. Produce a severity-ordered findings list; apply the fixes. If a structural fix needs real rewriting, loop back to the Writer (Phase 4) for that section, then re-run Phases 5–6 on the changed text ONLY.

Output: the publication-ready draft + an editor's findings list (what was fixed, what was cut, what looped back, promise-delivery verdict).

Phase 8 — Layout Designer (排版设计师) · Model: Sonnet

Job: Present the finished content with visual hierarchy, apply the mandatory citation format, and file it. Respect both Chinese and English publishing conventions.

1. Hierarchy: title, section headings, subheads where they earn their place.
2. Supporting detail: pull quotes, callouts, tables for comparative data, lists where prose would be dense, code/quote blocks.
3. Citations (MANDATORY): convert every inline citation marker to the final citation form and honor source precedence exactly as defined in conventions/quality.md (do not restate the format templates here — quality.md is the single source of truth for them).
4. CJK + English typography (apply directly, do not defer):
   • Chinese runs use full-width punctuation:，。：；？！「」（）— never ASCII ,.:;?!() inside CJK text.
   • Add a space between CJK and adjacent Latin letters or Arabic numerals (e.g. 用 Opus 写 not 用Opus写), unless the user's own voice sample consistently omits it.
   • Normalize mixed numbering (1. vs 1． vs (1)) to one style per the channel.
   • Keep heading levels and emphasis (**bold**) consistent throughout.
5. File it: write to writing/<slug>.md with frontmatter (title, type: writing, date, tags), citations in final form, and create back-links per quality.md (every mentioned person/company with a brain page gets a back-link from that entity's page).

Output: the laid-out, filed article at writing/<slug>.md, plus a one-line pointer to the slug it landed at.

Phase chaining & loop control

• Phases run in order, once each, by default.
• The only allowed loop is Editor (7) → Writer (4) for a section that needs real rewriting; after the loop, re-run Voice (5) and Fact-Check (6) on the changed text ONLY, not the whole piece.
• Never skip a phase. Never run two roles as one blended pass — the separation is the point.
• Human checkpoints (recommended, not mandatory): Phase 1 (topic/angle) and Phase 5 (voice) are the two decisions most worth a quick user confirm. Pause there if the request left them open; otherwise proceed and note the assumption.
• If a store is unavailable in Phase 2, log it and continue; degraded material/voice is acceptable, a blocked pipeline is not.

Output Format

The pipeline produces, in order:

1. Topic brief (Phase 1) — {working_title, thesis, audience, promise, spec:{length, depth, channel}}.
2. Material dossier (Phase 2) — (a) topic material + (b) voice samples, source-tagged.
3. Evidence map (Phase 3) — claim → [evidence], with UNSUPPORTED flags.
4. First draft (Phase 4) — complete prose sized to spec, title options, worked opening, citation markers.
5. Voiced draft + style fingerprint (Phase 5).
6. Fact-check report + corrected draft (Phase 6).
7. Editor findings + publication-ready draft + promise-delivery verdict (Phase 7).
8. Final article (Phase 8) — written to writing/<slug>.md with frontmatter, final-format citations, CJK typography applied, and back-links; the agent reports the brain slug it landed at.

The final brain page is the canonical deliverable. Intermediate artifacts are shown to the user as the pipeline runs; only the finished article is filed.

Anti-Patterns

• ❌ Running all eight roles on one model. They differ wildly in cognitive load; use the per-phase model routing table — Opus where judgment matters, cheaper models for mechanical passes.
• ❌ Re-scanning the stores in Phase 5 for voice samples. Phase 2 already gathered them in one pass; re-querying is wasted retrieval.
• ❌ Drafting (Phase 4) without the Phase 1 spec. Without length/depth/channel the writer guesses the size and gets it wrong.
• ❌ Collapsing roles into one pass ("write + fact-check + lay out at once"). The whole value is eight separate judgments; one blended pass does all of them badly.
• ❌ Inventing a voice in Phase 5 instead of learning it. Too little past writing → say so and ask for reference pieces; never fabricate a style and call it the user's.
• ❌ Letting an UNSUPPORTED or UNVERIFIABLE claim reach Phase 8. The fact-checker and editor exist to kill these; publishing one defeats the skill.
• ❌ Shipping ASCII punctuation inside Chinese text, or 用Opus写 with no CJK-Latin spacing. Phase 8 applies CJK typography directly — it is not optional polish.
• ❌ Skipping the citation-format conversion in Phase 8. The quality.md citation format is mandatory; inline markers are not the final form.
• ❌ Re-running the entire pipeline after a small Editor→Writer loop. Re-run only voice + fact-check on the changed text.

Tools Used

• query — brain-first material AND voice-sample gathering (Phase 2), evidence (Phase 3), prior-take checks (Phase 1).
• search — broader brain retrieval when query is too narrow.
• get_page — read specific prior articles in full for voice study (Phase 2).
• put_page — write the finished article to writing/<slug>.md (Phase 8).

External (non-gbrain) tooling, used best-effort via the session's own MCPs/CLIs:

• GetNote (mcp__getnote__recall*) and Obsidian (mcp__obsidian) — extend Phase 2 material and voice gathering.
• AI archive (ai-search) — prior cross-model conversations for Phase 2.
• Web fetch (firecrawl MCP or equivalent) — Phase 3 external evidence and Phase 6 re-verification of external claims.

The brain remains the primary and only required store; every external store is best-effort and the pipeline degrades gracefully when one is unavailable.

Related skills

• skills/article-enrichment/SKILL.md — enriches an EXISTING raw text dump in the brain; expert-writing CREATES a new article from an idea. Use enrichment when you already have the text; use this when you have only an idea.
• skills/strategic-reading/SKILL.md — reads source material into a playbook; can feed Phase 3 evidence but doesn't write the article.
• skills/concept-synthesis/SKILL.md — builds the intellectual map the topic strategist (Phase 1) and steward (Phase 2) can mine for angles and material.

Contract

This skill guarantees:

• Routing matches the canonical triggers in the frontmatter.
• All eight roles run as ordered, separated phases, each on its routed model; output is written under writing/ per writes_to:.
• The Phase 1 spec (length, depth, channel) is fixed before drafting and sizes the Phase 4 draft.
• The voice-polisher learns from the user's actual past writing (gathered once in Phase 2) before applying any style; it never fabricates a voice silently.
• Every external fact in the finished article carries a quality.md citation, traces to a source, and is re-verified by the fact-checker; UNSUPPORTED/UNVERIFIABLE claims do not reach publication.
• Phase 8 applies CJK + English typography directly and creates back-links per quality.md.
• Conventions referenced (quality.md, brain-first.md, model-routing.md, subagent-routing.md, _brain-filing-rules.md) are followed.
• Privacy contract preserved: no real names, no fork-specific filesystem path literals, no upstream-fork references in the skill itself.

The full behavior contract is documented in the phase sections above; this section exists for the conformance test.