Case study 01
An AI content pipeline that runs our own blog.
Four n8n workflows take a topic from first signal to a reviewed draft. We built it for ourselves, and it has been running since 27 January 2026.
The situation
Content is how our products get found.
Most people find SlideBazaar, SlideKit and SlideChef through search. For us, articles aren’t a side project. They’re how we get new users, and we have over a million registered.
Writing was never the slow part. The slow part was everything around it: picking topics worth covering, doing the research properly, and catching weak claims before a post went live. That depended on who had time that week, so we published in bursts and the quality varied.
What we built
Four workflows, one job each.
We built it in n8n instead of buying an AI writing tool. We wanted to own the workflow, swap models when better ones come out, and keep our data on our own systems. It has been running since 27 January 2026.
Stage 01
Content Listener
Keeps an eye on our sources and topics, and starts the pipeline when something comes up that’s worth writing about. Nothing runs on a timer.
Stage 02
Content Research
Uses Perplexity to gather sources and write a brief, so the writer starts with evidence instead of a blank page.
Stage 03
LLM Council
Several models read the same draft and argue about it. Weak claims and thin sections show up here, before anyone on the team spends time on it.
Stage 04
Article Writer
Writes the final draft in our house style, formatted and ready for someone to review, edit and publish.
The decisions that mattered
Three choices we’d make again.
01
Use more than one model
A single model can be confidently wrong in ways that are hard to spot. When several models review the same draft, the weak claims come out early and cheaply, before a person has to go through it.
02
Research before writing
Most AI content is bad because it’s written from nothing. Adding a proper research step before the writer made a bigger difference than any prompt changes we tried.
03
A person still signs off
Nothing goes live on its own. The pipeline gets a draft to where reviewing it is quick, but someone on the team still decides what gets published.
Results
Where it stands.
It has run every day since 27 January 2026. Every post still goes past a person before it’s published, and none have gone out without that check.
We haven’t been tracking output or cost per post properly yet. We’re adding that now and will put the real numbers here once we have a few months of them.
What we’d change
What didn’t go to plan.
Research matters more than writing, and we didn’t put enough into it at first. Tweaking the writer’s prompts helped a little. Giving it a better brief helped a lot more.
Which model we used turned out to be the decision that changed most often. We can swap a model in one node without touching anything else, and that has already saved us time more than once.
The problem we didn’t see coming was tone. Every model has its own voice, so a draft could pass review and still read like four different people wrote it. We fixed it with a single house-style guide that every stage writes to.
What this means for you
The same setup works for other kinds of content.
None of this is specific to blog posts. A trigger, a research step, a review step and a writer, with a person at the end, is the same structure we use for proposals, sales decks, reports and internal briefings.
We build it on your own licences (n8n, Make, Zoho Flow or Power Automate), write it all down, and hand it over so your team can change it without us.
Want this pipeline for your content?
We’ll build it on your tools, document it, and put the review step where it suits your team.