A one-person studio, with real money on the line

STILL BEING WRITTEN

Execution can be delegated. Judgment cannot. So I built a studio where agents execute and criticize, and money only moves on a human call.

Role: I built and operate the entire studio. The brand is my wife's business; the operation runs on systems I made.
Context: a small family e-commerce brand (pet fashion, print on demand), run outside working hours, with our own ad budget
System: an operations cockpit plus a plugin with 16 skills and 4 fixed-role agents
Period: 2026, in operation

The studio

The whole cycle runs on agents with fixed roles, like a real studio. Diagnosis reads the ad account and the funnel. Strategy prioritizes the test backlog. In production, an art director proposes visual routes anchored in reference, a producer executes the candidates, and an adversarial validator tries to kill every piece before I see it, with blocking criteria on two lenses: finish (execution, thumbnail legibility, correct copy) and semantics (does the piece have an idea?). I choose by looking at the roughs, never by reading a description. The approved campaign is born paused: turning it on is always a human gesture. Then measurement runs against fixed rules, and what the operation teaches gets promoted into versioned methodology, with semantic versioning and a changelog, like software.

Judgment, built in

Three mechanisms make the system trustworthy with real money. The validator is fail-closed: any blocking criterion kills the piece, so AI criticizes AI before a cent is spent. Campaigns are never born active: the system prepares everything, and turning it on is my decision with the numbers in front of me. And unit economics are the canonical rule: margins, CAC and breakeven ROAS are computed before any budget. No rule, no test.

A closed multi-agent loop: brief, art director, producer and an adversarial validator that can kill a piece, followed by a highlighted human decision gate, a campaign always created paused, measurement against unit economics, and learning promoted to a versioned method that feeds the next brief. — DIAGRAM · THE STUDIO LOOP AND THE HUMAN GATE

Results

Winning campaigns run at roughly 3 ROAS against a 2.4 breakeven.
An 80/20 strategy by niche, validated by data: broad audiences lost money, niche targeting profits.
Operational learnings promoted into versioned methodology: the studio improves every cycle, by contract.
Small and declared scale: a niche brand with our own budget. The case is not about size. It is about the system and the judgment, which hold at any scale.

(Space reserved for the 90-second video: brief, art director routes, candidates, the validator rejecting a piece, approved package, paused campaign.)