Music edits for tracks

Task

The system makes batches of short music edits to a 12–20 s excerpt of a track, for social media. All of them stay within shared frames of what looks good, yet differ in footage, finish and caption font — and none repeats past edits on other tracks. A Claude agent edits to the track’s score from a footage catalogue, Remotion renders, a check enforces the rules. What qualifies is decided by blind arenas.

How it works

1. Track → score: bars, hits, drop, mood
2. Lyrics: vocal stem → Whisper → caption screens
3. The system assigns each edit its footage, format, finish and caption style
4. Claude agent writes the edit to the score
5. Render — Remotion; check — rules and uniqueness across all past edits
6. Blind arena: only approved styles reach a batch

What had to be solved

The edits came out as copies

Left to choose on their own, the agents converged on one look: in an arena of six edits, four took the same caption style. Now the batch generator assigns each edit its own set — caption style and finish direction — and the check holds it. Uniqueness is measured against every past edit on every track, not only within the batch.

Commits

1. f6b8a3403.10.2026the batch generator assigns each edit its own set and diversity sheet: the copies in w2 came from leaving the choice to the agents
2. ce9c8d503.10.2026only approved styles and uniqueness across the whole history: every edit must differ from all past ones, not just its neighbours in the batch
3. 31950de02.10.2026ears: the agent is deaf — it gets the grid, hits, drop and mood as a score; the grid is checked against attacks instead of by ear
4. 5f266d006.10.2026first batch run the way the interface launches it: $2.06 for two edits — editing and finish by one agent cost less than the old finish alone
5. d463d5806.10.2026captions: word timing by the sound, screens on sixteenth notes and no shorter than 0.3 s