Annex — the AI animation studio system

Found by Frank on X, saved 02-08-2026. External material, kept whole as reference for the Make a film shelf. The revenue claims are the author's, not ours; the value is the production method.

The heart of it: the 4-block prompt architecture

Every prompt has the same structure. Three blocks are fixed forever; one changes per shot.

STYLE_LOCK    <- never changes - the medium
CONTINUITY    <- never changes - character design + world rules
DIRECTION     <- never changes - camera language + pacing
SHOT          <- the ONLY variable block per image

Forty shots across an episode; characters, world and style identical in every one. That is what makes it look like a show, not a collection of images.

Phase 0 — Style bible

One visual style, locked as a reusable style string appended to every prompt. Example: "2D cartoon, bold clean outlines, flat cel shading, saturated warm palette, slightly exaggerated proportions, soft shadow fills, consistent line weight throughout". Test until consistent; never improvise it per scene — drift starts the moment you freestyle the style.

Phase 1 — Character sheets (built once, used forever)

Two templates, in order:

Template A — base design (lock the design): style string + age/gender/build, skin tone, face, hair + OUTFIT BLOCK garment by garment + emblems ("original fictional designs, no resemblance to any real nation, brand, or regime") + body language + expression. "Full-body, front-facing, clean white background, flat studio lighting." Re-roll the same prompt, never rewrite it.

Template B — reference sheet (from the approved design): four views on neutral grey — full-body front, full-body rear, front close-up, 90° profile close-up. The OUTFIT BLOCK is pasted verbatim from A — never paraphrased. One word changed = visual drift across 40 shots. Any detail that must persist (a scar, a logo) is written into the sheet, not fixed across 40 scenes later.

Phase 2 — Location reference sheets

Four views of the same location, no characters, composed with deliberate negative space where a figure would stand. 2×2 grid, 16:9, labeled VIEW 1-4, consistent lighting and palette. Only the four VIEW lines change per location. A locked location is what makes episode 3 feel like the same world as episode 1.

Phase 3 — Voice design

One locked voice per character. Supplied to Seedance as a blank video with the audio track underneath — not an MP3 (the video carries the voice identity). One clip per line of dialogue; tag each line with @ to assign the voice. Two characters keep their own voices across one 15-second generation.

Phase 4 — Scene generation

Scene = character sheets + location sheet + voice refs + scene prompt. The scene brief is written as plain narrative with dialogue verbatim, then an agent converts it to the multi-shot cinematic prompt:

[start - end] shot size, focal length, camera angle,
camera move - character description, action, expression:
"dialogue" @VoiceRef

The pace IS the format: 2-3 s per shot · ~6 shots per 15-second generation · action/reaction alternating · wide → medium → close-up rotation · focal length cycled 24/35/50/85 mm (the rotation is what makes it feel edited) · camera moves explicit, never ambiguous · sound cues called per shot · hard limit 1,500 characters per prompt (the Seedance limit). If a voice drifts across a multi-shot: single-shot generations, cut in edit.

The range trick

The same character sheet works across genres: the same girl in a soft slice-of-life scene and a horror battle — only the STYLE LOCK palette line and DIRECTION block shift; CONTINUITY stays identical. One asset library serves bedtime stories and action series.

The automation loop

A standing agent (the author uses Higgsfield Supercomputer scheduled tasks): every Monday it pulls the week's script, breaks it into ≤15 s scenes, writes every shot prompt from the sheets, outputs one production file, and flags missing character/location sheets BEFORE credits are spent. Planning in free mode; renders flagged separately.

The score (Suno prompt shapes)

Structure a music prompt like a brief: instruments leading, what enters when ("full orchestra enters at 0:20"), the emotional arc, runtime, loop behavior, two artist references for the energy, register ("memorable but not aggressive"). Commercial license on the paid plan.

The author's four rules

1. Assets before scenes — never generate a scene with a character that has no sheet.
2. One style string, everywhere.
3. Explicit beats implicit — persistent details live in the sheet.
4. The library is the show — episodes are recombinations of library assets; the asset library is what you are really building.