[Case study / 2025]
The blog system
Nine agents, one signature
A writing system that turns a one-line premise into a published, sourced article: specialized agents for research, drafting, fact-grounding and a de-AI editing pass, then CMS publishing and a social queue. Every post still ships under one human byline.
- Year
- 2025
- Team
- Luis Alejandro Martínez
- Deliverables
- Multi-agent drafting pipeline · Claim verification · De-AI editing pass · CMS publishing · Social scheduling
Problem
Writing regularly is a second job on top of the first one. Research, drafting, sourcing, images, meta descriptions, publishing and then the social queue are seven separate chores, and the last four are the ones that quietly stop happening.
The obvious answer is to hand the whole thing to a model, and that is exactly how a blog loses its credibility. Generated posts invent prices, model names and product states that were never true; they carry the rhythm of a machine; and a reader who notices once stops trusting the byline. I wanted the volume without the slop, and the byline is mine.
Challenge
Facts go stale on a schedule. Prices change, models get superseded, products ship or die, and a paragraph that was right in June is wrong in August. So the pipeline cannot only ask a model to write well; it has to fetch the current state of the world first and then hold the writer to it. A separate research step produces a pack of claims with source URLs, and the body gets checked against that pack before it is allowed to publish, with a small, explicit list of things the writer is forbidden to invent.
The second problem is voice. Models default to a register that reads as generated regardless of the language: the tricolon, the rhetorical question, the summary paragraph nobody needed. The fix is a dedicated editing pass, but that pass is dangerous: an editor that can add a sentence can add a claim. So the pass is constrained. It may not add or drop a claim, and its output is thrown away and the original kept if it touches a media placeholder, unbalances a code fence, or shrinks the post below sixty percent of its length. When a voice profile is set, the profile wins, so a writer who uses em dashes keeps them.
The last constraint is cost and time. Nine agents is nine chances to spend tokens on work nobody reads, so each post type only runs the agents that earn their place, and freshness research runs only where facts are load-bearing.
Product
Everything starts as a markdown file with YAML frontmatter: a premise, a post type, a target audience, keywords, links and any media I want placed. The post type picks the writer. General posts get the SEO specialist, technical posts get the engineer, narrative posts get the storyteller, product reviews get the reviewer, and market analysis gets the data analyst. The same field picks the freshness mode: research for general and technical, which fetches the input links and the analysis sources and writes a claim pack that the writer must respect; links only for reviews; off for narrative.
After the draft comes the editing pass, which strips the tells of machine writing under the constraints above and cites Wikipedia's catalogue of those tells as its rulebook. Then the pipeline generates a hero image and any inline illustrations, and replaces the media placeholders with real components: YouTube, SoundCloud, galleries, SVG diagrams, PDFs, placed where the text actually calls for them rather than in a block at the end.
Publishing is the part that has to be boring. Posts go to the CMS with their hero image uploaded to the CDN, their categories created if missing, and their draft or published state tracked so a re-run updates instead of duplicating. From there the same content feeds a social queue that schedules the post across six platforms for the week. The whole thing is TypeScript run through the project's Make targets, so a post is four commands: create the template, generate, upload as draft, publish.
Results
The pipeline is what the blog runs on. Each post below was a premise and a set of links before it became a page with images, sources and a social queue behind it, and I still read every one before it publishes because the byline is mine.
- 100×
- monthly searchimpressions
- 20 min
- from premiseto published post
- 33
- posts publishedthrough the pipeline
A post takes about twenty minutes from premise to page when I publish it straight away; scheduled posts wait for their slot. Each one is grounded in a claim pack fetched that day rather than in whatever the model remembered. That is the part worth having: the volume is automated and the checking is not.
Search is where it shows. In July 2025, the month the pipeline started, the blog's posts drew 17 impressions on Google; in September 2026 they drew 1,708 and 40 clicks, and October held at 1,478 impressions. Clicks are still small, and that is the honest next step: letting search data pick the topics the pipeline chases, so readers who see a post also open it.
Next
A problem
like this one?