· Adriano Martins Barcellos

A one-person studio, run on 52 AI skills

By the fourth blind round, its writing beat posts made with a professional template system in 5 of 6 verdicts.

The problem

One person, several brands, live clients. Each brand needs its own voice, content every week, and nothing published that isn't true. By hand, that doesn't scale. Handed to a generic AI tool, everything starts to sound the same.

What I built

A system of 10 plugins and 52 skills, running on Claude Code. Every brand has a profile: its voice, its visual rules, and what it must never say. The skills cover strategy, content, client reporting and search, plus a senior-review pass that pressure-tests a plan before it ships.

Client work runs in its own lane. My writing method only runs on my own brands; a client brand runs in client mode, where its own rules lead. That is enforced in code, with 24 tests, not left to good intentions.

One run, start to finish

A real run from October 2026: one of my own log posts turned into a five-slide carousel. The system wrote the slides, checked them against the schema and rendered them in the brand. No photos were generated for this run. Select any image to see it full size.

The brief: source post, brand, angle, audience, proof points and call to action
1 · The brief goes in
The five slide roles (cover, hook, core, application, call to action) with each headline
2 · The structure: one job per slide
The five rendered slides, from 'Every brand I run kept a diary. The studio holding them didn't.' to 'Read the full piece on the log'
3 · Rendered in the brand, ready to post
Schema validation passing, then the render with zero overflow on all five slides
4 · Schema and overflow checks, run before anyone sees it

How I know it works

I tested the writing blind against posts made with a professional template system, which had then been edited by hand. In round one, judged by me, the template system was better on every count. From then on, an AI judging panel scored each pair from three reader roles (a social media editor, a local reader, an event marketer), with the labels hidden and the order randomised. By round four, the engine won 5 of 6 verdicts, and the one loss was by a single point.

What stood out: the template system's posts had passed through four human editing steps. Mine had none. The engine now has its own editor pass built in to do that job; it hasn't been re-tested blind since.

These were my own tests, not an independent study. The method is written down, and the blind files are kept.

What I'd do next

A recorded walkthrough of a live run, a blind re-test with the editor pass in place, and an AI video ad test with real hook and click numbers.