Agent Atlas

The Builders

PJ Ace

Runs Genre.ai, an AI-native ad studio, and publishes the filmmaking workflow behind his narrative tests.

Cut the script first, lock the looks in image grids, then hand those frames to the video model.

Steal this

  • Condense the script before you generate anything. He cut a 20-minute audiobook passage to 3 minutes.
  • Lock each character's look in still images first. The video model follows the frames.
  • Expect volume: one battle took more than 100 clips of 15 seconds.

The lesson and the takeaways are our summary. Everything below is theirs, with where they said it.

The setup

Factual
Machines

Not stated.

Software

An LLM for the dialogue cut, Luma's agent for image grids, Nano Banana Pro for the frames, and Seedance 2.0 for the shots.

Models

Luma's agent, Nano Banana Pro and Seedance 2.0. No coding-agent model named.

As of March 18, 2026. Setups change; follow the links.

How they work

Factual
  1. On 18 Jul 2025 he said his team made a 25-minute, fully AI TV show in 3 months. (PJ Ace)
  2. On 18 Mar 2026 he condensed a 20-minute audiobook passage to a 3-minute video. An LLM stripped out the dialogue, then he cut it again. (PJ Ace)
  3. He gave Luma's agent the passage, his script and shot list, and had it make 30 options for each character's look. (PJ Ace)
  4. Reference images of the characters in the location keep the location consistent from shot to shot. (PJ Ace)
  5. The frames go into Seedance 2.0. One battle was more than 100 clips of 15 seconds. (PJ Ace)