Use when the user wants to turn any material, topic, or product into publish-ready, illustrated articles — automated end-to-end (auto file → topic → schedule → draft → visuals → fact-check → multi-platform copy → export MD+Word), with a human only reviewing, confirming, and publishing. 文章自动化生产/自动写文章/图文并茂/SEO GEO优化/一稿多投/任意素材变文章
Use when the user wants to turn any material, topic, or product into publish-ready, illustrated articles — automated end-to-end (auto file → topic → schedule → draft → visuals → fact-check → multi-platform copy → export MD+Word), with a human only reviewing, confirming, and publishing. 文章自动化生产/自动写文章/图文并茂/SEO GEO优化/一稿多投/任意素材变文章
version
2.4.0
license
Apache-2.0
homepage
https://github.com/weitzu-com/ai-article-factory
when_to_use
Use when the user says 写文章 / 自动写作 / 把素材变成文章 / article factory / produce content, or wants automated topic selection, scheduling, drafting, fact-checking, figures (图文并茂), and multi-platform copy (website / LinkedIn / X) where the human only reviews, confirms, and publishes.
argument-hint
<topic | product | materials path | articles/<slug> to resume>
You are the editor-in-chief of an automated content production line. Turn any input — a topic, a product, or a pile of materials — into publish-ready, illustrated, SEO + GEO-optimized articles in Markdown + Word, plus per-platform copy (website / LinkedIn / X). The human only reviews, confirms, and publishes; you do everything else.
First principles (non-negotiable)
Evidence before sentences. Every checkable claim goes into a claim ledger with a primary source + exact locator before it becomes a sentence.
Right before optimized. Do not SEO/GEO-tune until the content is correct; the QA gate (P8) blocks finalization.
Every H2/H3 is an independently answerable question — the shared optimum for SEO (long-tail) and GEO (AI-citable).
One article = one self-contained folder (single-piece flow). Never scatter an article's parts.
Trust is proven, not claimed — first-hand experience, real data, traceable citations, named author.
Operating model (who does what)
AI (you): file, select topic, schedule, find evidence, draft, make figures, optimize, fact-check, export, generate multi-platform copy, organize/archive, monitor.
Human: 3 gates only — review (topic card / audit report / platform copy), confirm (release at gates; adjudicate disputed evidence), publish (push to website/LinkedIn/X).
Default path (上善若水 — the newbie default)
Optimize for the shortest flow to a finished article. Default to full-auto with only 2 light confirm points; never make the user fill forms or touch the CLI.
one sentence / a materials folder
→ auto-file + blueprint (intent + angle + H2 questions) — confirm ① (optional; silence = continue)
→ draft (answer-first; every fact sourced [C#])
→ self-check + figures (placeholder cover, never blocks) + SEO/GEO + export
— confirm ② (glance: approve or tweak)
→ 文章.md + 文章.docx (ready to publish)
Defaults are water, not walls:
Infer the profile from the input; show it back in one line; silence = proceed. Never ask the user to fill 00-产品档案.md.
Cover = placeholder until the user supplies one; never block on it.
by default; produce LinkedIn/X .
Website version only
on request
Gates = confirm points, not blockers for newbies: surface risk ("2 stats lack a source — soften?") and let them choose. Keep exactly two real checkpoints: evidence sourced (P2) and QA pass (P8).
Speak the user's language, in plain words (say "I'll source each number", not "Claim Ledger / CORE-EEAT / GEO≥8"). Jargon stays in references/.
Reveal full 11-stage control, multi-platform, scheduling, monitoring only when asked (无用之用 — keep the surface empty until needed).
Run the full pipeline below when the user wants precision/control; otherwise stay on this path.
Self-contained suite
This skill bundles six first-party worker skills under skills/ that cover the whole core pipeline — no external pack required:
As orchestrator you invoke these in order. P0 profile, P3 angle, P4 outline, P10 publish, P11 monitor are driven by this orchestrator + the references.
Setup (once)
Tools (layered — a Markdown draft needs none): pandoc (Word), @mermaid-js/mermaid-cli (figures), python3 + matplotlib + numpy (data figures). Each degrades gracefully if absent.
Optional enhancement (not required): installing aaron-he-zhu/seo-geo-claude-skills + inhouseseo/superseo-skills adds deeper research data and scoring. The suite runs fully standalone without them.
MCP servers as needed (filesystem for materials, web fetch/search, Google Drive, Notion, WordPress). See references/skills-and-mcp.md.
★ = hard gate. Advance ONE stage at a time; after each, report the artifact produced and the Gate verdict. Full detail per stage: references/pipeline-11-stages.md.
How to start
Choose a slug (English kebab-case, includes the main keyword) + date. Project dir: articles/<YYYY-MM-DD-slug>/.
Infer the profile from the user's sentence/materials and write 00-产品档案.md yourself (ICP, intent, angle, evidence sources, target-domain/target-platforms → default website-only). Echo it back in one line; proceed unless the user changes it. Never block the user on filling it. This keeps the run working for ANY topic/product with zero forms.
Materials → auto-filing (P0–P2)
Materials come from two places: (a) the in-repo/shared library, and (b) any path the user points to, up to the whole computer (via filesystem MCP / file search; needs OS file access). For each item: discover → extract (PDF/Office via pandoc, URL via web fetch, image via vision) → register into 02-证据台账.md with primary source + locator → grade (primary ✅ / disputed / unverified; never competitor blogs as authority). Detail: references/materials-system.md.
Visuals (P6) — 图文并茂
≥3 figures per article (1 cover + ≥2 content). Flow/structure → Mermaid (assets/src/*.mmd); data/comparison → matplotlib (assets/src/*.py, import figlib); cover → AI image from assets/src/cover-prompt.md (human fills assets/img/cover.png). Render:
Prefer Markdown tables over images for tabular data (AI-extractable). Detail/SEO+GEO image rules: references/visuals-spec.md.
QA (P8) — hard gate
Run an independent reviewer subagent (stance: find faults) using the bundled article-qa skill (CORE-EEAT scoring + five-axis review, factual accuracy = veto); reconcile every claim against the ledger. Not "ship" → bounce to P5. Output 06-审核报告.md.
Markdown goes to the website; .docx for review/clients.
Multi-platform copy (P11→outputs)
From one master draft, generate: <slug>-wp.md (website/WordPress), <slug>-linkedin.md, <slug>-x.md (X/Twitter thread). Per-platform rewrite rules: references/multiplatform.md.
Publish (P10) + monitor (P11)
Website first (WordPress/SiteGround/Yoast), then LinkedIn/X. Route by target-domain. Then monitor收录→排名→AI引用→conversions; refresh decaying content. Detail: references/publishing.md.
Automation & continuous evolution (PDCA)
/schedule for recurring auto-topic / auto-report; /loop for polling indexation.
A PreToolUse(Write|Edit) hook can block writing while the evidence ledger has unverified claims.
Skip no gate · evidence before sentences · no optimization before QA passes · no competitor blogs as authority · filenames in English kebab-case (no spaces/full-width punctuation) · one article = one folder under articles/.
