Script in.
Film out.
Motion graphics generated entirely from code, for coding agents: no video-generation models. A small plan names the beats; the engine solves the layout with real fonts, times it to reading speed, moves it with a theme and joins the scenes with transitions.
Work in progress · v0 · not the finished siteSame plan, two themes
One 7-scene plan rendered twice. The theme owns colour, type and motion: editorial reveals by line and stays calm; bold slams word by word. The scenes are joined by carry (a shared element travels), the theme's wipe, push and cut.
Known in these renders: scenes still clear before the next one starts, so some joins sit near-empty, and long holds are static. An independent review measured both; the timing rework that overlaps outgoing and incoming scenes is being built now.
How a film is made
- Script
- Plan (one LLM call)
- Validate + re-plan
- Solve layout, real fonts
- Time to reading speed
- Theme motion + transitions
- Render (HyperFrames)
- Gates
The plan holds only beats, on-screen copy, stress words and the intent of each join (continue, shift, break). No timings, coordinates or eases: everything after the plan is deterministic, so the same plan and theme always render the same film. Copy that does not fit is never shrunk below its floor; the repair ladder tries the next layout, one type step, then a split.
Every layout, every theme, every aspect
The still matrix renders each layout at rest for 3 beats × 2 layouts × 2 themes × 3 aspects × 5 copy fixtures (short, longest, one long word, Cyrillic, tall). A cell either fits inside the safe zone or reports why, with numbers. Grey cells are copy the layout refuses by design.

Gates an agent runs on its own video
The same kit checks finished films: frozen stretches, labels outside the safe area, labels colliding, loudness, encoding. Tested on 11 finished lesson reels; an independent reviewer labelled samples of the findings by looking at the frames.
| Round | Findings that were real | What changed |
|---|---|---|
| e1 | 20% | Generic checks: mostly noise (stacked caption layers, cross-fades, ad-sized safe zones). |
| e2 | 73% | Per-reel contract, region-based still detection, a checker inside the animated figure itself. |
| e2c | 90% | Figure clock bug fixed, same-label collisions counted, drift-split holds merged. |
Samples of 46, 34 and 30 findings. Precision only: what the gates miss is not measured yet.
Next
- Overlapping transitions and staged holds (in progress)
- A playable studio: every option as a control, live preview, render button
- 12-film proof: 6 scripts × 2 themes against a naive baseline, judged blind
- Then wide expansion: more beats, themes, inserts from the HyperFrames catalog, voice-over timing