What If Animation

What If animation

One topic line in, a finished 15-25 second episode out. Every frame is SVG that Claude writes (rendered by Remotion); the voice and the sound come from ElevenLabs. Two examples ship in studio/src/episodes/, both ready to render: rewind ("What if you could rewind 10 seconds?", the proof clip: shots played backwards as the power) and copy-paste ("What if you could copy and paste anything?", the skill's first full run: a new kitchen set, a schematic for the science beat, a cliffhanger blackout). Read the one closest to your topic before you start.

Rules

• Original characters only. Pip (the hero, look B) and Mo (his friend) are ours, and so are the sister, mum and the suit on the rig. Never a channel's own hero, and nothing that resembles a famous cartoon character: check every new look on the cast sheet (Mo's first hat was dropped for drifting toward one).
• Original stories. The formula is shared; the episode isn't anyone else's. Don't remake a channel's title.
• Never claim the viral channels are made with Claude or any other tool. Nobody has said how they're made.
• Keys by name only. ELEVENLABS_API_KEY, from the environment or a .env above the studio. Never print, paste or hardcode a key.
• Report the ElevenLabs credits of every run: the scripts log them to out/<slug>/costs.json and master.py totals them in report.json. The voice's cost is read from the account's history (exact); sound clips don't show there, so their cost comes from the account counter, which can lag: re-read it before you report. Say the number when you hand the episode over. The copy-paste run cost 152 credits (78 for the voice, 74 for 12 new clips).
• Storyboard before animation. No video render until the stills pass the self-check (step 5) and, when the person is in the loop, they've seen the board.
• Every fact is real. A number in the script is worked out and sourced in the episode's science field first.

Setup (once per machine)

Needs Node 18+, ffmpeg and ffprobe on the PATH, Python 3.9+ with numpy and Pillow, and an ElevenLabs key.

bashCopy code
cp -r <skill>/studio ./what-if-studio && cd what-if-studio && npm install

Everything below runs from that studio folder; <skill> is this skill's folder. To work offline or in your own brand face, put a display font in public/fonts/ and set titleFont in src/studio.config.ts (the default title face, Bricolage Grotesque, and the UI faces load from Google Fonts at render time). To update an existing studio to a newer engine, copy <skill>/studio/src/engine over its src/engine.

The pipeline

Work through the steps in order. Each ends on something you look at before the next.

1. Script

Read references/formula.md. Search YouTube for the premise first: if a What If channel already has that exact episode, change the power or the rule. From the topic line, decide the power's word, its rule, three escalating demos and the turn (one real fact with a number). Write src/episodes/<slug>/episode.json with id (PascalCase, the composition name), slug, topic, title (two lines), script (one entry per beat) and science (each fact: the claim, the working, the source). 45 to 65 words. Read it aloud at pace: every line must be a picture.

2. Voice

bashCopy code
python <skill>/scripts/voice.py <slug>

One ElevenLabs take with character timestamps (the 356-character copy-paste take cost 78 credits), then the pauses cut to the genre's pace: public/<slug>/voice.wav and src/episodes/<slug>/words.json. Check the length (15-25 s) and the word times it prints. Retune pauses with gaps in episode.json and re-run: the take is cached, so that's free. Change a word and it calls the API again.

3. Shot list

Design one shot per picture in the script (usually 9 to 14), varying the size and angle every cut (wide, insert, medium, close-up, over the shoulder, high, low; see references/look.md). Add to episode.json:

• shots: { "id": "S1", "size": "medium", "set": "kitchen", "key": 14, "note": "what we see" } (key is the frame of the shot's own clock the storyboard shows)
• edit: { "shot": "S1", "at": "cookie" } per cut, cued on the word it lands on (plan.py lists the cue forms; reverse, from, to and look for powers that bend a shot's clock)
• fx: one-frame flashes and shakes, { "at": "S3+2", "kind": "flash" }

bashCopy code
python <skill>/scripts/plan.py <slug>

It writes timeline.json and out/<slug>/shotlist.md and prints the pace. Aim for 27+ cuts a minute in the opening and no static shot over about 3 s; fix what it flags before drawing anything.

4. Storyboard

Read references/kit.md for the building blocks. Write src/episodes/<slug>/shots.tsx, one component per shot, each a pure function of its clock t (copy the patterns in the two examples' shots.tsx), and index.tsx (the episode and its power look), then add the episode to src/episodes/index.ts. At this step each shot only needs to be right at its key frame. When a topic needs a place the library lacks, build it with the room kit and add it to src/engine/sets/ (and to the skill's copy) so the next episode has it, as the copy-paste run did with the kitchen and the schematic. A new character or costume: add the look, then check it on the cast sheet (npx remotion still src/index.ts Cast out/cast.png).

bashCopy code
npx tsc --noEmit
node <skill>/scripts/stills.mjs <slug>
python <skill>/scripts/board.py <slug>

out/<slug>/storyboard.jpg: one captioned still per shot.

5. Self-check the stills

Look at the storyboard, then open every still in out/<slug>/board/ at full size, and run the storyboard checklist in references/look.md honestly, shot by shot. Write down what fails, fix it, re-render just those shots (stills.mjs <slug> S3 S7), rebuild the sheet and look again. Repeat until it passes. Check references/traps.md for the failure you're seeing before inventing a fix. If the person is in the loop, show them the board now and take their notes before animating.

6. Animate

Bring each shot to life from its key pose: poseAt keys with anticipation, holds and overshoot, faceAt for expressions, camera moves with kf, the power's look, the fx. Then render a draft and check the motion:

bashCopy code
python <skill>/scripts/master.py <slug>
python <skill>/scripts/motiongrid.py <slug> S2 S5
node <skill>/scripts/stills.mjs <slug> --frames 120 240

Read the grids like a flipbook against the motion checklist in references/look.md; fix and re-render until it reads.

7. Sound

Read references/sound.md. Add sound to episode.json: the clips (prompt and length), one event per visible action cued to its shot, ambiences per location, the music bed and the levels. Check what the library already has, then generate:

bashCopy code
python <skill>/scripts/sfx.py <slug> --dry
python <skill>/scripts/sfx.py <slug>

It prints the credits it used and sets the stem volumes in timeline.json. Then check that every word stays clear:

bashCopy code
python <skill>/scripts/mixcheck.py <slug>

It lists the moments, while a word is being spoken, where effects or music come within 3 dB of the voice. A hit that illustrates a word lands in the pause just after it, never on it; a bed (a hum, a drone) under a key line comes down; if most of the stem crowds the voice, lower sfx_lu. Re-run sfx.py and the check until only deliberate overlaps remain (a title boom, a turn).

8. Master and hand over

bashCopy code
python <skill>/scripts/master.py <slug>

Renders, brings the mix to -14 LUFS with peaks under -1 dBTP, and writes out/<slug>/report.json (length, loudness, pace, credits). Watch the final file once through with sound. Hand over out/<slug>/<slug>-final.mp4 and out/<slug>/storyboard.jpg, with the length, the ElevenLabs credits used, and anything you'd still improve.

Where an episode lives