Building an AI Content Pipeline From Idea to Published Post in 30 Minutes
So I built a pipeline last quarter that takes an idea — literally a sentence, like “the best AI tools for HR teams” — and produces a published WordPress post 30 minutes later, fully researched, written in my voice, internally linked, SEO-optimized, and image-illustrated. I wasn’t trying to put bloggers out of business. I was trying to ship a content site without becoming a content team. The pipeline I built is the same architecture that the system you’re reading runs on, so I have skin in this. Here’s how it works and how you’d build your own. This isn’t a sales pitch — it’s an open recipe.
The stack
You need five things:
- **n8n** (self-hosted or cloud) — the orchestration layer. ($0 self-hosted; $24+/month cloud)
- **OpenRouter** — single API for all the LLM calls. (~$5-$50 per pipeline run depending on word count and models)
- **Tavily** — research API for live web search. ($30-$100/month at moderate volume)
- **Google Sheets** — data layer (article queue, outputs, style examples). (free with Google Workspace)
- **WordPress + WP All Import** — publish target. (free with hosting; ~$200 one-time for the plugin)
Total stack: roughly $50-$200/month at moderate volume, plus per-pipeline run AI costs that work out to $0.30-$0.80 per published article.
The workflow
1. Intake. A new row in the Google Sheet `Queue` tab with the idea, primary keyword, secondary keywords, and article type. Status starts as `new`. The n8n workflow watches for new rows and picks them up.
2. Research planner. A small LLM call (Gemini Flash, cheap) reads the idea and generates 5-8 specific research questions. Output: JSON list of questions.
3. Tavily research. For each question, hit Tavily’s search API with `includeanswer=true, includerawcontent=true, maxresults=5`. Collect the sources.
4. Research synthesis. A bigger LLM call (Claude or GPT) reads all the Tavily results and produces a structured brief: facts (with sources), stats (with year + source), quotes (with speaker + source), entities, counterpoints. Output: JSON.
5. Approval gate 1. The pipeline writes the research brief to the Sheet and changes status to `awaitingresearchapproval`. You read it, edit if needed, mark approved. Pipeline resumes.
6. Outline. Claude builds an H2/H3 outline from the brief, with per-section word budgets and intent. You review and approve.
7. Drafter. Claude writes the full first draft from the outline + brief + 2-3 sampled style examples + must-follow style rules. Output: markdown.
8. Style match. Another Claude call rewrites the draft to match the style fingerprint of your existing examples — sentence-length distribution, opener patterns, transition phrases.
9. Humanizer. Two passes. First sweep removes obvious AI tells (overused transitions, em-dash abuse, vague intros). Second pass: the model rates itself for AI-likelihood and rewrites the 5 worst sentences.
10. Linker. Pull internal links from a daily-cached `/sitemap.xml`. Pull external links from the research brief sources. Insert deterministically, validating external URLs with HTTP HEAD.
11. SEO + Schema + Image (parallel). Generate meta title (≤60 chars), meta description (≤155), slug, focus keyword. Generate JSON-LD schema. Pick categories and tags from your existing taxonomy. Generate an image prompt and call DALL-E (or Replicate, or Ideogram). All four happen in parallel.
12. QA editor. Final structured audit — factual pass, structural pass, style rule scores, link counts, banned phrases, word count. If `overall=fail`, corrective edit and re-QA. Max 2 retries.
13. Approval gate 3. Status to `awaitingfinalapproval`. You read the article, mark approved. Pipeline writes to the `WP_Ready` tab.
14. WordPress import. WP All Import polls the Sheet on a schedule, ingests new rows, creates posts in draft status. You hit publish.
The 30 minutes is total wall-clock from “new row in queue” to “draft in WordPress,” with you reviewing at the three approval gates. The autonomous AI work is about 8-12 minutes; the rest is your review time.
What this saves
A competent freelance writer charges $200-$500 per researched, SEO-optimized blog post and takes 4-8 hours. This pipeline produces something within striking distance at roughly $1 of AI cost per article and 30 minutes of your time. For a content site shipping 4 articles a week, you save $40K+ per year and reclaim 100+ hours.
What it doesn’t save you: the strategy work. Picking the right ideas, defining the audience, building the style fingerprint, doing the topic research that points to which keywords to chase — that’s still you.
Gotchas
Style fingerprint depth matters. Garbage examples produce garbage output. Spend the time on 4-10 strong style examples that show range.
Hallucination risk on research. Always validate stat claims and source URLs. The QA pass catches most issues but not all. Treat the output as a polished first draft, not a final manuscript.
Approval discipline. The temptation is to approve gates without reading. Don’t. The pipeline produces better work when you actually catch outline issues at gate 2 instead of fixing them at gate 3.
Cost discipline. Per-pipeline costs creep up with bigger models. Tune the model assignments — use cheap models for planning, linking, schema; use premium models only for drafter, style match, and QA. Watch the OpenRouter dashboard.
FAQ
Is this just Writer Ultimate?
Writer Ultimate is one implementation of this pattern. Other working implementations include LangChain’s content workflows, custom n8n setups, and Zapier flows for simpler versions. The pattern is the value, not the specific tooling.
Can I publish without human review?
Technically yes. Practically no. Even with 95% quality output, the 5% problem articles will tank your SEO and your reputation. Keep humans in the loop.
What if I don’t use WordPress?
The publish step is the swappable part. Ghost, Webflow, Notion-as-CMS, and custom backends all work. The pipeline writes structured data; whatever consumes it gets the same input.
How long until I can build this myself?
If you’re technical with n8n already: a weekend. If you’re starting from zero: two to three weeks of evenings. The free Writer Ultimate setup is a fast track.