A 40-node content pipeline that runs without supervision
Long-form content was produced manually, with someone watching each run. A 40-node n8n pipeline now researches, writes, validates and publishes unattended three times a week, at roughly ten cents of model cost per article.
The client
US medical-device manufacturer. Published anonymously at the client’s request — no company name and no person’s name appears in this entry.
The problem
Producing long-form content meant a person driving each step and checking the output before anything went out.
That puts a ceiling on output equal to one person's attention: the work could only run as often as someone had time to supervise it.
The risks that make unsupervised writing dangerous all needed answering before supervision could be removed — two runs claiming the same topic, a duplicate of something already published, a confidently invented statistic, a link to a page that does not exist.
What was built
- A production n8n pipeline of 40 nodes using the Claude API, with Supabase as its store.
- Research, drafting, validation and publishing in one unattended run.
- Keyword locking: a topic is claimed before work starts, so two runs can never produce the same article.
- Duplicate detection before any model call — the cheapest place to stop a wasted run.
- A 22-check quality gate covering word count, heading structure, internal links and keyword density, plus detection of fabricated statistics and invented links.
- Cost gates that stop the expensive steps once an earlier one has failed.
- Automatic retry, and a weekly pass that refreshes ageing posts.
Before and after
| Before | After |
|---|
| Supervision | A person watched every run | Unattended |
|---|
| Cadence | As often as someone had time | Three scheduled runs a week |
|---|
| Model cost per article | Not tracked | About $0.10 |
|---|
| Checks before publish | One manual read-through | 22 automated checks |
|---|
| Duplicate topics | Possible whenever two efforts overlapped | Prevented before work starts |
|---|
How each figure was measured
- Supervision
- Whether a person is required for a run to complete.
- Cadence
- The pipeline's schedule.
- Model cost per article
- Model API spend per completed article, as reported by the operator. Infrastructure and any human review are not included in that figure.
- Checks before publish
- Count of checks in the quality gate. They include fabricated-statistic and invented-link detection, which is what makes an unsupervised run defensible at all.
- Duplicate topics
- A topic is locked at the start of a run, so a second run cannot claim it.
What is deliberately not here
No traffic or ranking figures appear here. This entry is about the pipeline that produces the content; what the content went on to do is a separate measurement nobody has closed out yet.
The other engagements
- Eleven data sources in one dashboard — and stale numbers labelled stale
- A seven-day reporting cycle became a scheduled evening job
- Eight contract sources consolidated — and the silent data loss stopped