Running four departments as one person — building a video production team out of AI agents

I wanted to publish video consistently. I am the only person here. The way out of that contradiction was to give the AI four departments.

The problem

The goal was a steady output of vertical short-form product videos (1080×1920). Each video needs four stages — planning, production, quality checking, analysis — and running all four by hand every time is far too much work for one person.

A one-off is a matter of effort. Several a week, indefinitely, is not: something gives eventually. And what gives first is the checking. When you are tired, QA gets generous, and weak videos go out. “The person tries harder” has two ceilings built into it — a limit on volume, and drift in quality.

The diagnosis

Breaking the process apart, very little of it actually required me. Research, production, scoring and aggregation can all be executed by machine once the criteria are written down. The only parts genuinely mine were approving the handoff between stages and the final publish.

Put differently, the real fault was structural: one person playing all four roles. The quality gatekeeper was the same person as the producer, which is why the scoring had gone soft. The fix was to separate the roles and give each an independent pass criterion.

What was built

A four-department AI team in Claude Cowork — planning, production, QA, analysis.

• Planning — researches product candidates and drafts the video structure
• Production — automated rendering via a Python script. Gradient backgrounds and review-distribution bar charts generated in code; narration via TTS, with output cached by SHA-256 so nothing is regenerated twice
• QA — scores against a 100-point rubric and returns anything below threshold to production automatically. Because the scorer is a different role from the producer, in-house leniency disappears structurally
• Analysis — daily cron on GitHub Actions pulls the statistics automatically

The essential part is the pass gate between departments. Work does not advance unless the previous stage’s output meets the standard, and the person only approves the passage through the gate.

The outcome

• Videos scoring 96/100 produced with the human doing nothing but approving and publishing
• The entire pipeline persists as code and documentation, so nothing depends on one person. Hand it to a different AI model tomorrow and it continues at the same quality
• Separating QA into an independent role structurally removed the “the maker grades their own work” problem

What this means for your business

The common way to fail with AI is to build one all-purpose assistant and hand it everything. Separate the roles, put a pass line between the stages, and quality stabilises while the person concentrates on the final judgement alone.

Any work shaped like “plan → produce → check → review” can use the same structure. It does not have to be video. Where the human holds the decision and where the AI takes over — that boundary is part of what I design and hand to you.