AI-Assisted Article Pipeline for Claude Code
Raw idea in, fact-checked article out. Every claim traceable. Every stage human-confirmed.
A Claude Code skill that transforms raw ideas into professional articles through a 5-stage semi-automated pipeline — with built-in source verification, cross-reference checking, and anti-AI-voice writing principles.
English | 简体中文
Most AI writing tools optimize for speed. Content Alchemy optimizes for trust.
You: "Write an article about AI replacing jobs."
AI: "According to a McKinsey report, 800 million jobs will be displaced by 2030..."
You: (checks) That number is from 2017 and the actual range was 400-800M with heavy caveats. 😤
AI: "Studies show that 47% of US jobs are at risk..."
You: (checks) That's the 2013 Frey & Osborne paper, widely criticized for methodological issues. 😤😤
Stage 2: Sources extracted → each one gets an authenticity check
Stage 3: Source Truth Table built → every claim mapped to its origin
Stage 3.5: Cross-Reference → same claim checked against 2+ independent sources
Stage 5: Confidence self-assessment → weak claims downgraded or flagged
[unverified]Result: An article where readers can trust the data. ✅
| Problem | Solution |
|---|---|
| AI makes up data | Source Truth Table forces citations |
| Source exists but data is wrong | Stage 3.5 Cross-Reference adds multi-source verification |
| Article reads like AI | 7 anti-AI writing principles plus a 6-dimension scan |
| Uncertain facts slip through | Confidence self-assessment downgrades or marks weak claims |
┌─────────┐ ┌─────────┐ ┌─────────┐ ┌─────────┐ ┌─────────┐ ┌─────────┐
│ Stage 1 │───►│ Stage 2 │───►│ Stage 3 │───►│Stage 3.5│───►│ Stage 4 │───►│ Stage 5 │
│ Topic │ │ Source │ │ Deep │ │ Cross- │ │Refining │ │Humanized│
│ Mining │ │Extraction│ │Analysis │ │Reference│ │ │ │ Article │
└─────────┘ └─────────┘ └─────────┘ └─────────┘ └─────────┘ └─────────┘
│ │ │ │ │ │
▼ ▼ ▼ ▼ ▼ ▼
mining- Source Source Cross-Ref manifesto.md article.md
report.md Authenticity Truth Table Report
Report
Every stage pauses for human confirmation before proceeding. No auto-pilot.
git clone https://github.com/AliceLJY/content-alchemy.git
claude skill add ./content-alchemyalchemy AI and loneliness # Full pipeline from topic
alchemy-setup # Check/download dependencies
Start from any stage:
| Mode | When to use | Entry point |
|---|---|---|
| Topic Mode | Start from scratch | alchemy [topic] → Stage 1 |
| Source Mode | Have transcripts or articles | Provide sources → Stage 3 |
| Draft Mode | article.md already exists |
Hand off to content-publisher |
Content Alchemy is the research & writing half. Publishing is a separate repo.
content-alchemy (this repo) content-publisher
───────────────────────── ──────────────────
Stage 1: Topic Mining Image generation
Stage 2: Source Extraction ───► Layout formatting
Stage 3: Deep Analysis WeChat publishing
Stage 3.5: Cross-Reference Cleanup
Stage 4: Refining
Stage 5: Humanized Article
| Repo | Role |
|---|---|
| content-alchemy (this repo) | Research and writing, Stages 1-5 |
| content-publisher | Images, layout, publishing, cleanup |
| recallnest | Shared memory layer for Claude Code, Codex, Gemini CLI |
| openclaw-tunnel | Docker-to-host tunnel for /cc, /codex, /gemini |
| digital-clone-skill | Build digital clones from corpus data |
| cc-shell | Lightweight Claude Code chat UI |
| telegram-ai-bridge | Telegram bots for Claude, Codex, and Gemini |
| telegram-cli-bridge | Telegram CLI bridge for Gemini CLI |
| Version | Era | Highlights |
|---|---|---|
| v1.0-v2.5 | Antigravity | Project inception, 9-stage to 7-stage workflow evolution |
| v3.1-v3.2 | Claude Code (Opus 4.5) | Bug fixes, fallback mechanisms, doc restructuring |
| v4.0-v4.3 | Claude Code (Opus 4.6) | Chrome reuse, Bilibili extraction, Cross-Reference, 6-dimension AI scan |
| v5.0 | Claude Code (Opus 4.6) | Split into content-alchemy and content-publisher, progressive reference loading |
Project Structure
content-alchemy/
├── README.md
├── README_CN.md
├── SKILL.md
├── references/
│ ├── stage1-mining.md
│ ├── stage2-extraction.md
│ ├── stage3-analysis.md
│ ├── stage3.5-crossref.md
│ ├── stage4-refining.md
│ ├── stage5-writing.md
│ ├── source-channels.md
│ └── video-acquisition.md
├── scripts/
│ ├── format-text.ts
│ ├── preprocess_article.py
│ └── sync-codex-skill.sh
├── templates/
│ └── cross-reference-report.md
└── assets/
└── wechat_qr.jpg
Prerequisites & Environment
Required:
- Claude Code
- Bun
- Python 3
- A writable
~/Desktop/bot-articles/
Local path assumptions (may need adjustment):
- Output directory:
~/Desktop/bot-articles/{topic-slug}/ - Persona source:
~/.claude/projects/*/memory/writing-persona.md - Style catalog:
~/.openclaw-antigravity/workspace/images/style-catalog.md - Style history:
~/.openclaw-antigravity/workspace/images/style-history.txt
Compatibility:
- Tested on macOS with the author's Claude Code CLI workflow
- Not guaranteed to work on Linux or Windows
- Some paths are hardcoded — users may need to adjust persona and style asset paths
- Downstream publishing depends on
content-publisher
Built by 小试AI (@AliceLJY) for the WeChat public account 我的AI小木屋.
Six content pillars: Hands-on AI, AI Pitfall Diaries, AI and Humanity, AI Cold Eye, AI Musings, and AI Visual Notes.
MIT