WeChat Article Pipeline Example¶
Goal¶
Use Hermes Agent to write, detect AI-taste, de-AI-fy, and publish WeChat Official Account articles. The skill is continuously optimized using real account analytics data — not guesswork.
Maturity Level¶
L2 (semi-autonomous). Agent drafts the full article and generates diagrams. Human reviews and publishes. The analytics feedback loop runs weekly to refine title strategy, content structure, and topic selection.
Files¶
examples/wechat-article-pipeline/
├── SKILL.md # wechat-viral-article skill (main pipeline)
├── scripts/
│ ├── ai_detect.py # 8-dimension AI-taste scoring engine
│ └── de_ai.py # AI-taste removal + repair instructions
└── references/
└── article-template.md # Obsidian YAML frontmatter template
How It Works¶
The 9-Step Writing SOP¶
Step 1: Topic selection → 3 directions (source code / pitfalls / comparison)
Step 2: Research → web search + official docs + GitHub issues
Step 3: Outline → choose from 4 proven templates
Step 4: Draft → 2000-4000 characters, inline image markers
Step 5: AI-taste detect → scripts/ai_detect.py, target <20%
Step 6: De-AI-fy → scripts/de_ai.py + 8-step manual method
Step 7: Generate diagrams → draw.io CLI, brand color system
Step 8: Final check → headline checklist + layout + image positions
Step 9: Save to Obsidian → /今日头条/{title}.md with YAML frontmatter
The Analytics Feedback Loop¶
The skill includes real account data from July 2026 (13 followers, 8 articles, 238 recommendation reads/week) that validates:
- 「XXEngineering」naming convention is the strongest IP asset — Engineering-series articles averaged 81 reads vs 57 reads for non-series articles
- Concept/framework articles outperform tool/list articles — frameworks averaged 108 reads, tools averaged 55
- Recommendation algorithm favors high-completion + high-engagement content — publish every 2-3 days during the boost window
These findings are baked into the skill's hard requirements checklist:
- Title format: XXEngineering: subtitle (mandatory)
- Content type: concept/framework/methodology only (no tool lists)
- Minimum 1 architecture diagram + 2 code screenshots
- Ending: interactive question + next-preview teaser
- AI concentration: <20%
Loop¶
- Read
STATE.mdto check the publishing schedule and next planned topic. - Consult the performance data of past articles to select the next topic with highest potential.
- Agent writes the full draft following the 9-step SOP.
- Run
scripts/ai_detect.pyon the draft. If concentration >20%, runscripts/de_ai.py. - Generate architecture diagrams via draw.io CLI (brand colors: #1a1a2e dark, #00d4ff blue).
- Save final article to Obsidian vault + paste into WeChat backend.
- Update
STATE.mdwith publish date, actual reads after 48h, and lessons learned.
What Makes This Different¶
Most article-writing pipelines just generate text. This pipeline:
- Detects and removes AI-taste programmatically (8-dimension rule engine + LLM-assisted rewrite)
- Is data-driven — the skill file itself contains real analytics from the account's first month, so every writing decision is informed by what actually worked
- Bakes the IP into the naming convention — the account's "XXEngineering" series identity is enforced at the title level
- Generates professional diagrams via draw.io with a consistent brand color system
Dependencies¶
- Hermes Agent v0.6+
- draw.io desktop app
- Obsidian vault at
/Volumes/DISCa/Hermes Vault/Hermes Vault - Python 3.10+ (for ai_detect.py and de_ai.py)