wechat-article-writer

从选题到发布的公众号一体化写作工作流

Multi-agent pipeline: Orchestrator delegates writing and reviewing to independent subagents. The orchestrator never writes or reviews — it routes, tracks versions, and enforces quality gates.

Setup

bash <skill-dir>/scripts/setup.sh <workspace-dir>

Installs: bun runtime, bundled baoyu renderer deps, and a persistent preview server (wechat-preview.service, port 8898, auto-restart).

Scope

Handles: Topic research → Chinese-first writing → quality review → scrapbook illustrations → WeChat formatting → publishing to WeChat draft box (via Official Account API or CDP browser automation).

Does NOT handle: Git/version control, non-WeChat platforms, post-publish analytics, WeChat messaging/customer service.

Ends at: Article saved to WeChat draft box. User publishes manually.

---

Commands

Trigger any command below, or see skill.yml for the full trigger pattern list.

Command	What it does
forge topic X	Research trending angles, propose 3 options with hooks
forge write X	Full pipeline: research → publish (9 steps)
forge draft X	Write + format only, stop before illustrations/publish (steps 1-7)
forge publish <slug>	Publish an existing draft to WeChat
forge preview <slug>	Render preview, run format quality checks
forge voice train	Analyze past articles to extract voice profile
forge status	Show pipeline status and pending drafts

If no subject given, loads from session.json (set by forge topic). See references/data-layout.md.

---

Pipeline (9 Steps)

State persists to pipeline-state.json — survives compaction. See references/pipeline-state.md.

#	Step	Who	Details
1	Research + Prep	Orchestrator	(a) web_search for topic angles + 5-8 sources. (b) Verify each source exists (fetch title/authors/venue). Save as sources.json. (c) Load voice profile. (d) Generate outline (6-8 sections), save outline.md.
2	Write	Writer subagent	Chinese-first draft. Writer MUST cite only from sources.json or mark [UNVERIFIED]. See references/writer-prompt.md
3	Review	Reviewer subagent	Blind 8-dimension craft scoring. See references/reviewer-rubric.md
4a	Revise (auto)	Writer subagent	Max 2 automated cycles. Loop back to Step 3 if score < threshold.
4b	Revise (human)	Human-in-the-loop	If still below threshold after 2 auto cycles, user provides direction. Pauses pipeline.
5	Fact-check	Fact-Checker subagent	Verify every claim via web search. Produces corrections + reference list. Max 2 fact-check cycles (corrections → re-verify). See references/fact-checker-prompt.md
6	Format	Script	bash scripts/format.sh <draft-dir> [draft-file] [theme] — baoyu renderer (default theme: classic WeChat style). Themes: default/grace/simple. If fact-check required >3 text changes, Orchestrator does a spot re-review (Reviewer scores only changed paragraphs, not full article).
7	Preview	Human	Open http://<host>:8898/formatted.html (persistent preview server, systemd wechat-preview.service), await text approval
8	Illustrate + Embed	article-illustrator + script	Generate scrapbook images (AFTER text approval). ~$0.06/article via Z.AI (preferred, ~$0.015/image) or ~$0.50 via OpenRouter.
9	Publish	Orchestrator	Three paths — check in order: (C) WeChat Official Account API via appid+appsecret (credentials at wechat_secrets_path in config.json) — preferred, most reliable; (A) OpenClaw browser tool with base64 chunking for macOS/Titan; or (B) direct CDP WebSocket for Linux/remote. Paths A+B use two-phase injection (text first, then images via clipboard blob paste). See references/browser-automation.md

Key Rules

• Writer never self-reviews. Reviewer is blind — never sees outline or brief.
• Illustrations LAST. Most expensive step. Only after user approves text.
• article-illustrator is the ONLY image method. Must follow full scrapbook pipeline: read references/scrapbook-prompt.md → generate JSON plan with 300-500 char descriptions → call generate.py. Never bare prompts. Prefer Z.AI provider ($0.015/image, 97.9% Chinese text accuracy) over OpenRouter ($0.12/image).
• Two-phase image injection. Base64 images are stripped on save. Inject text-only HTML first, then insert each image at the correct position via clipboard blob paste (WeChat auto-uploads to CDN). Verify image count + positions after insertion.
• Browser tool vs direct CDP. On macOS/Titan where OpenClaw manages the browser, you MUST use the browser tool (Path A). Playwright isolates page contexts — external CDP connections see zero targets. On Linux with standalone Chrome, use direct CDP (Path B). See references/browser-automation.md.
• Base64 chunking for browser tool. Raw HTML in the browser tool's fn parameter breaks due to escaping conflicts. Always base64-encode HTML, store in chunks via window._b, then atob() and inject. Track chunks_stored in pipeline state for compaction recovery.
• Always save as draft. User publishes manually.
• Check for WeChat API credentials first. If wechat_secrets_path credentials file (see config.json) exists, use Path C (API) — no browser required, more reliable. Fall back to Path A/B only if no credentials.
• ensure_ascii=False is mandatory for WeChat API. requests(..., json=payload) escapes Chinese as \u5199\u4e66. Always use data=json.dumps(..., ensure_ascii=False).encode('utf-8').
• Topic fidelity: Every revision preserves the article's 初心 (purpose statement in pipeline-state.json). Drift = FAIL.

Image Counts by Type

Type	Min	Max
科普	3	5
教程	3	6
观点	2	4
资讯	2	3

---

Review Dimensions

Reviewer scores 0-10 on craft-observable dimensions (not outcome predictions):

Dimension	Weight
Insight Density (洞察密度)	20%
Originality (新鲜感)	15%
Emotional Resonance (情感共鸣)	15%
Completion Power (完读力)	15%
Voice (语感)	10%
Evidence (论据)	10%
Content Timeliness (内容时效性)	10%
Title (标题)	5%

Pass: weighted_total ≥ 9.0, no dimension below 7, Originality ≥ 8.

Hard blockers (instant FAIL): 教材腔, 翻译腔, 鸡汤腔, 灌水, 模板化, 标题党.

Full rubric with scoring criteria: references/reviewer-rubric.md

---

Architecture

Orchestrator (Main Agent) — routes, tracks, enforces gates
    ├── Writer Subagent — drafts + revises (Opus model)
    ├── Reviewer Subagent — blind scoring (Sonnet model)
    ├── Fact-Checker Subagent — verifies claims via web search (Sonnet model)
    └── article-illustrator — scrapbook images (after text passes)

---

Configuration

Configure via ~/.wechat-article-writer/config.json (generated by scripts/setup.sh):

Field	Default	Description
default_article_type	"教程"	Default article type (科普/教程/观点/资讯)
wechat_secrets_path	~/.wechat-article-writer/secrets.json	Path to WeChat API credentials
chrome_debug_port	18800	Chrome CDP port for browser automation (Path B)
wechat_author	—	Author name shown in WeChat draft
word_count_targets	See defaults	Min/max word counts per article type

See references/data-layout.md for full config schema.

---

References

File	When to load
references/writer-prompt.md	Step 2 (writing) and Step 4 (revision)
references/reviewer-rubric.md	Step 3 (review) — full 8-dimension scoring criteria
references/fact-checker-prompt.md	Step 5 — claim extraction, verification, correction protocol
references/viral-article-traits.md	Step 2 — Writer self-check list
references/pipeline-state.md	On resume or compaction — state machine schema + protocol
references/browser-automation.md	Step 9 — Two publishing paths: Path A (OpenClaw browser tool) and Path B (direct CDP). Includes base64 chunking, image insertion, save verification.
references/LESSONS_LEARNED.md	Hard-won lessons from production publishing sessions (escaping, selectors, mixed content, costs)
references/data-layout.md	Directory structure, slug generation, config/session schemas
references/agent-config.md	Setup — Gateway, AGENTS.md, environment config
references/quality-checks.md	Steps 3, 7 — content/format quality gates
references/figure-generation-guide.md	Step 8 — illustration placement heuristics
references/wechat-html-rules.md	Step 6 — what HTML/CSS works in WeChat
references/templates.md	Step 1 — starting templates by article type
references/voice-profile-schema.json	Step 1 — voice profile field definitions
references/default-voice-profile.json	Step 1 — fallback voice profile