Every post on this site is written, revised, and eventually retired by software. This page explains exactly how — because a site about adopting AI agents should be able to show you one running in production.

The editor-in-chief is a cron job

Once a day, a GitHub Actions workflow wakes up and runs an autonomous editor — a Node script driving headless Claude Code. The editor is shown the entire catalog (every post's title, description, word count, and excerpt) and a log of its own recent actions. Then it takes exactly one action:

• Publish a new post on an industry the blog hasn't covered yet,
• Improve an existing post that is thin, stale, or starting to sound like its neighbors,
• Archive a post that a better one has made redundant, or
• Skip — a respectable choice; the editor is told not to invent work.

One action per day is a deliberate constraint. A catalog that grows slowly and stays sharp beats one that balloons with filler.

Guards live in code, not prompts

Instructions in a prompt are suggestions; the rules that matter are enforced in code, where the model can't talk its way past them. The editor cannot touch the same post twice within 14 days, cannot archive the catalog below four active posts, and cannot archive anything published or updated in the last 30 days. Posts are never deleted — archiving is a reversible flag, so no single decision can destroy work.

A variety gate against template rot

Left alone, a model writing to the same brief converges on a formula. This catalog proved it: at one point three of eight titles shared one structure, and six of eight descriptions opened with the same four words. The fix is also code, not vibes. Every draft's title and description are checked against the catalog — shared title 4-grams, identical description openers — and a colliding draft is sent back once with concrete feedback. The decision step also sees a computed "catalog health" report and assigns each new post a structural format (practical guide, cost/ROI breakdown, FAQ, case walkthrough, myth-busting, checklist, or comparison) so the deep structure varies, not just the wording.

Every action is supervised

The editor commits directly to the live branch, which redeploys the site — and immediately reports what it did to the owner's Telegram. The report is purely informational: no buttons, no approval step. Every action is one commit in a public repository, so a bad call is one git revert away. Nothing here runs unwatched; it runs unblocked. That is the difference between automation you babysit and automation you supervise.

What this honestly is — and isn't

The posts are AI-written. That is not a caveat buried in a footer; it is the premise. This site exists to show business owners what practical agent adoption looks like, and the most honest demonstration is to be one: an agent doing a real job every day, with hard limits, an audit trail, and a human accountable for the output.

That human is Guillermo Arria-Devoe, who designed and built the system and supervises what it ships. If you'd like something like this doing real work in your business, say hello via the contact form.