Stickman Animation Studio — /make-video
Turn one prompt into a finished, studio-quality English vertical cartoon (1080×1920, 30 fps, H.264 + AAC) using nothing but code. You write the story, storyboard it with an explicit narration-to-visual sync contract, run the bundled engine through its QA gates, and hand back the verified MP4 with its sources.
Operate autonomously. No confirmation questions, no progress narration, no stopping halfway. If a step is blocked, treat it as a routing problem: try the next tool, the next asset, or build it yourself. Only stop when the video is rendered, verified, and delivered.
Production language is English — exclusively. Narration, dialogue, on-screen text, captions,
labels, and insert copy are all English. The narration must sound like professional documentary,
educational, or animated storytelling (references/10-english-voice-direction.md).
0. Probe your capabilities, then adapt
Before producing, inspect what THIS agent/model can actually do, and adapt the workflow. Do not assume; check. One minute here saves a broken render later.
| Probe | How | If missing → adaptation |
|---|---|---|
| Shell + file access | run node --version | No shell at all → this skill cannot render; fall back to install/PASTE-THIS-PROMPT.md handoff to a coding agent |
| Web research | a web-search tool available? | No search → write from internal knowledge, mark every claim "unverified" in fact-check.md, prefer evergreen physics/history topics |
| Vision on images | can you read image files? | No vision → rely on engine/scripts/qa_stills.py + engine/scripts/qa_sync.ts pixel/lexicon gates and say so in the delivery report |
| Background execution | can you run long commands in the background? | No → run renders in the foreground and use --preview first; budget for several minutes |
| Context budget | very large context? | Tight context → read only the references/ files the current stage names, one at a time; keep the story JSON as the single source of truth |
| Reasoning depth | strong model → push craft further: richer staging, more beats, deeper research | Weaker model → follow the reference defaults literally; they are calibrated to produce good results without improvisation |
Whatever the environment, the quality gates (steps 4, 5, 7) are not optional — a weaker workflow must still pass the same executable checks.
0.5 Resolve the engine
The engine is a self-contained Remotion project. Find it in this order:
- A directory in the current project whose
package.jsonis namedstickman-studio-engine— use it. - The
engine/folder that ships next to this file — copy it into the workspace (cp -r <skill-dir>/engine ./engine) and runnpm installinside it (first time only). - If neither exists, the skill was installed without its engine — clone the skill repository
and take
engine/from there.
One-time setup inside the engine directory:
npm install # Remotion + three + zod
pip install edge-tts imageio-ffmpeg numpy scipy pillow # narration + procedural audio + QA
The 65 SFX, 6 music beds and film grain are generated procedurally on first render (no downloads,
no licenses, no third-party audio). Narration needs internet for edge-tts (free, no key);
--no-tts works fully offline for staging work. English voices only — see
references/10-english-voice-direction.md for the voice catalog and direction rules.
1. Research the topic (English)
- Input is a topic, a title, or a path to a TXT file (read it if given).
- Research with web search before writing. Build a fact-check table (claim | what the source says
| source URL) for every factual claim you will narrate →
stories/<slug>/fact-check.md. - If something is only rumored or folklore, either drop it or mark it in-narration ("legend has it…").
- Prefer the weirdest true detail over a smoother invented one.
2. Write the script + choose the voice
Aim for 190–210 words ≈ 60–75 s. Structure (follow it exactly):
- Line 1: paradox hook stated as plain fact. No greetings, no "Did you know".
- Setup and stakes — hero, place, want.
- Escalation + a short quoted line where someone mocks the hero.
- Turning point / rhetorical question.
- "You see," + the mechanism, with an everyday analogy and an insert graphic.
- Twist on the last line. End hard.
Full craft rules: references/04-script-craft.md. Save the script to stories/<slug>/script.md.
Then direct the voice before rendering: pick provider/voice/rate for the genre from
references/10-english-voice-direction.md (default: en-US-AndrewMultilingualNeural @ +22%;
calm documentary reads slower). Test the hook line with a quick synthesis if unsure.
3. Check and build assets
- List what the story needs: places, standing objects, held items.
- Check what exists:
npx tsx scripts/locations.ts --list. Reuse before building; never make a near-duplicate; name assets for what they are. - Build missing assets before storyboarding — recipes in
references/08-extending-assets.md(3D location / 3D object / 2D prop / 2D background / outfit). - Verify new 3D assets with
npx tsx scripts/locations.ts --stills --only=<name>.
4. Storyboard WITH a sync contract (the sync gate)
Write stories/<kebab-title>.json. Schema + ids: references/09-schema-reference.md,
engine/src/catalog.ts. Craft rules: references/05-storyboard-craft.md.
Narration-to-visual synchronization is mandatory and machine-checked. Every narration segment
(a shot's say) must visibly show what it says: right place, right time of day, the described
action as a pose/beat/prop, mentioned objects on screen, quoted lines bound to a quote actor,
and continuity across shots. Word-level timing inside a shot is bound with beats (at = a
spoken word).
Run the executable gate and iterate to zero errors:
npx tsx scripts/qa_sync.ts stories/<file>.json # place/time/action/object/dialogue/continuity checks
npx tsx scripts/qa_sync.ts stories/<file>.json --map # print the shot-by-shot sync contract table
- All ERRORs must be fixed (change the shot's world/time/pose/props, or rephrase the narration).
- Review every WARN; either fix it or record why it is acceptable in
stories/<slug>/sync-map.md(generate the table with--map; template:templates/sync-map.md). - A story-specific lexicon (idioms, invented nouns) goes in
stories/<slug>/sync-lexicon.json— see the template.
5. Stills gate (never skip)
npm run make -- stories/<file>.json --stills # one QA still per shot → out/<slug>/stills/
python scripts/qa_stills.py out/<slug>/stills # exposure/dead-frame gate (exit 1 on fail)
Then inspect every still for framing, faces, floating props, readability, consistency — with your
own vision if you have it, else agentvision check <file> per still plus the pixel gate above.
Fix and re-run until clean. Never render video before this gate passes.
6. Render
npm run make -- stories/<file>.json # full quality → out/<slug>/<slug>.mp4
npm run make -- stories/<file>.json --preview # half-size draft first when unsure
Run in the background when possible; do not start unrelated heavy jobs concurrently. Narration is cached per slug and re-synthesizes only when script/voice/rate change.
7. Verify (ffprobe + frames + audio gate)
ffprobe -v error -select_streams v:0 -show_entries stream=width,height,r_frame_rate,codec_name -of csv=p=0 out/<slug>/<slug>.mp4
ffmpeg -i out/<slug>/<slug>.mp4 -vf fps=1/5 qa_%03d.jpg # pull frames across the video
python scripts/qa_audio.py out/<slug>/<slug>.mp4 # loudness/clipping/voice-presence gate
- Specs: 1080×1920 (540×960 for
--preview), 30 fps, h264 + aac; duration ≈ narration + 0.6 s. - Frames: first/middle/last not black; captions legible; no stuck shot.
- Audio gate must PASS: −14 LUFS ±1.5, true peak ≤ −1.0 dBTP, voice energy present.
8. Repair loop, then deliver
Loop stages 4→7 until every gate passes. Budget: up to 3 repair rounds per gate; if a gate still fails after that, deliver with an explicit note about what remains (never silently ship a known failure, never stop without a delivery).
Deliver as a set: the MP4, script.md, fact-check.md, sync-map.md, the story JSON, any new
library assets, and honest limits (verified fact vs. dramatized; which gates ran and how they
were checked).
Quality bar (non-negotiable)
- Sync: the visuals show exactly what the narration communicates — zero contradictions, zero dead shots. When narration and animation would tell two different stories, rewrite the storyboard.
- Expressive characters: reactions on beats, poses that read in silhouette, one consistent light source per scene; idle life (breathing/blink/walk) is automatic — intentional motion on top.
- Cinematography: framing ladder per scene, camera variety (the sync gate flags ≥3 identical framing+camera shots in a row), one clear visual purpose per shot.
- Sound never fights the voice: ducking is automatic; manual
sfxonly for specific moments; the audio gate enforces loudness and no clipping. - No repetition, no slop: vary framings, transitions and reactions; ≤ 2 fx per shot; ≤ 1 insert per 8–10 shots.
- No AI-generated images or video, no image-generation services, no paid APIs. Ever.
- English only — narration, dialogue, on-screen text, captions.
Deep reference material (read per stage, not all at once)
| File | Read it for |
|---|---|
references/01-animation-principles.md | motion, easing, anticipation, secondary action — as the engine implements them |
references/02-character-system.md | the stickman rig: proportions, builds, poses, faces, consistency |
references/03-scenes-camera-light.md | parallax layers, framing, camera moves, lighting, locked palettes |
references/04-script-craft.md | script structure, hooks, analogies, word budgets |
references/05-storyboard-craft.md | shot grammar, beats, transitions, the sync contract |
references/06-sound-design.md | ducking, auto-SFX map, music moods, comfort rules, audio gates |
references/07-rendering-qa.md | all QA gates, pixel checks, ffprobe, repair loops |
references/08-extending-assets.md | recipes: 3D location/object, 2D prop/background, outfit, sfx |
references/09-schema-reference.md | every story JSON field, with defaults and constraints |
references/10-english-voice-direction.md | the English voice catalog, direction, pacing, offline fallback |