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.

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.

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

Review Dimensions

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

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):

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

References