# aimator by Auditmos

A CLI that walks one person through producing a short animated episode.

Eleven stages, each consuming only files earlier stages produced, never conversation context. One project can carry a complete set of assets on two image tracks, which yields two independent animations from one screenplay. Polish narration by default.

## How it is made

Eleven stages in the three phases every production knows. Each stage gets only the files approved in the one before it: a check verifies and writes nothing, and only an explicit approval records that a person said yes, bound to the exact bytes, so editing a file outside the tool revokes it.

Pre-production, shared by both versions of the film: 0 Preparation (the idea, the episode source file and character photos become the series rules, the cast and the episode's decisions), 1 Screenplay (scenes with timings, dialogue and sound), 2 Character (a character sheet, eight views and a reference portrait, once per image track), 3 Shot list (scenes broken into shots and clips), 4 Prompt package (a description of every future frame and clip).

Shoot, once per image track: 5 References (places, props and faces seen more than once), 6 Opening frame (the film's first frame, the only one drawn from words alone), 7 Clips (a video shot with its entry and end frames, each starting where the last one ended).

Post-production, once per image track: 8 Edit (the silent picture cut), 9 Narration (the voice over the picture, the one version with nothing under it), 10 Music and effects (the film with everything in it).

## Changelog

### 0.1.0 · 2026-09-21

The first episode went the whole way: from an idea in a file to a film with sound.

The same screenplay was drawn by two image models. Each track has its own references, its own opening frame and its own clips, so one idea in a file yields two films you can compare side by side.

Episode: Dzielna Ewa · 01 — Burza · 16:9 · 1:30 · PL.

- Eleven stages run end to end: stage 0 gathers the rules and the cast, stage 10 hands back a film with a soundtrack.
- Two image tracks run independently from the episode's first picture to the finished film.
- Every artifact carries its own acceptance, bound to bytes: check verifies and writes nothing, approve accepts, and editing a file outside the tool lapses that acceptance.
- The film comes back in three cuts — silent episode.mp4, narrated.mp4 with the narration, mixed.mp4 with everything — and each carries its own yes.
- Every paid command says how many calls it will make before it makes them, and has a dry run that prints the full prompt without reaching for a key.

These notes describe the repository at commit b29a344. The material is one episode — “Burza” from the series “Dzielna Ewa” — produced between 17 and 20 September 2026; work in progress since then is not included. The films here are transcoded to H.264 so that any browser plays them: the originals are HEVC 1920 × 1080, 60 MB on the gpt-image track and 79 MB on seedream. The pictures are JPEG previews scaled to 1600 px; the originals are PNG, 2816 × 1584 for frames and 1920 × 1920 for character cards.

Source baseline: [b29a344](https://github.com/auditmos/aimator/commit/b29a344692a0a1a5528e95d3780b09bf40424c3e) (private repository; access required).

Episode source file: [source.md](/sources/0.1.0/source.md). The file an episode starts from: the idea, what the episode refuses to do, and nothing else. Stage 0 copied it into the workspace and recorded its sha256, so everything below is traceable to these 703 bytes.

- gpt-image: [Film](/media/0.1.0/gpt-image-mixed.mp4) (1:30, 1920 × 1080, 8.8 MB), 9 stills.
- seedream: [Film](/media/0.1.0/seedream-mixed.mp4) (1:30, 1920 × 1080, 11.1 MB), 9 stills.

- Stage 1 (screenplay.md): [screenplay.md](/documents/0.1.0/screenplay.md). Scenes with timings, written in Polish from the source file. One text-model call for the whole film.
- Stage 3 (shot-list.md): [shot-list.md](/documents/0.1.0/shot-list.md). The screenplay broken into shots and clips: what is on screen, how long it runs and what is heard. The edit is derived from this rather than from a second file holding the same truth.
- Stage 9 (narration.md): [narration.md](/documents/0.1.0/narration.md). The sentences the narrator reads, lifted from the shot list rather than written anew. The validator finds every sentence inside the shot it names.
- Stage 10 (sound-design.md): [sound-design.md](/documents/0.1.0/sound-design.md). The cue sheet: what plays, from which second, and under which shots. A cue's declared shots must be exactly the shots its seconds cover.

[View this release](/releases/0.1.0/)