Claymation ad

A clay character has a problem. The product fixes it. Twenty-six seconds, vertical, cut like a real broadcast spot.

You write the beat map yourself, in conversation, and show it to the user before anything is spent. The scripts do the mechanical half.

What makes this format, and not just "an animated video"

Two things, and if you drop either one it stops reading as a commercial.

1. The cut rhythm. A story film cuts to the narration — each shot holds as long as its sentence takes to say. An ad is the opposite: the rhythm is fixed first and the voice has to fit inside it. Six movements, measured off a real broadcast spot:

 movement   shots   length                      what it does
 ────────────────────────────────────────────────────────────────────
 setup        6     1.2s → 0.3s, accelerating   the problem piles up
 beat         1     4.9s, one long hold         the character lands it
 montage     15     0.07s – 0.5s                the transformation
 payoff       1     3.1s                        calm after the noise
 hero         1     1.5s+                       the pack, alone
 endcard      1     2.5s                        brand + tagline

The acceleration into the montage and the sudden stillness after it are the whole feel. Even cuts throughout read as a slideshow.

2. Two looks, intercut. Clay characters, then a photoreal macro insert of the real material — suds on a plate, fibres in a cloth, the sponge going in. Clay alone looks like a cartoon; macro alone looks like stock footage. Cutting between them is what makes a clay world feel like it has real dirt in it.

"looks": {
  "clay":  "Stop-motion claymation shot on a real miniature set: …",
  "macro": "Macro product photography, real materials, shallow depth of field, …"
}

Each shot names one. The setup and montage lean macro; the beat and payoff are clay; the hero is the pack itself.

Before you start

cd skills/claymation-ad
cp .env.example .env     # then fill in the keys for the backend you chose

The inReels column is tested, not aspirational — the whole pipeline has been run through it. It replaces two provider accounts with one billed in credits, but not quite all: the vision check needs an Anthropic key whichever backend paints the frames. SOUND_PROVIDER=inreels is newer than the rest: the call is verified against production, but no complete ad has run through it. Drop that line and keep an ElevenLabs key if it gives you trouble. Without CLAUDE_API_KEY step 2 paints without checking its own work and says so, rather than failing after you have paid for the frame.

INREELS_API_KEY=ir_live_…      # inreels.ai/dashboard/settings/api-keys
IMAGE_PROVIDER=inreels
MOTION_PROVIDER=inreels
VOICE_PROVIDER=inreels
SOUND_PROVIDER=inreels
INREELS_MOTION_MODEL=seedance-2.0-mini-frames   # 3 credits per 5s

For this format in particular, Seedance mini is the one to reach for: it is about a quarter the price of Kling and holds comic timing better, which is what a clay performance lives on. On inReels voices are named slugs (allison-millennial), not ElevenLabs ids — list_voices gives the forty.

ffmpeg is required either way.

Method

1. Write the beat map — you, not a script

The user gives you a product. Write work/ad.json yourself.

• Time every shot by hand. seconds on each one. This is the format; do

not leave it to be derived.

• The problem must be visible in shot 1. Not a logo, not a title — the

mess.

• One joke, landed on the beat shot. That single long hold is the only

place a performance fits.

• The montage is materials, not story. Suds, fibre, water, shine. Nobody

reads a montage cut; they feel it.

• Every character gets a locked `look` — face, ears, nose, clothing

colours. Clay faces carry broad expressions well and subtle ones badly, so write big emotions.

• Describe what IS, never what is not. "The plate is clean and shining"

works; "the plate is not dirty any more" paints a dirty plate.

• Write `product.pack` properly — colours, shape, what is printed on it.

It is painted once and used as a reference everywhere the pack appears.

• The voice fits the picture, not the reverse. ~2.6 words a second, and

the end card is not talked over.

Then check the rhythm in code, because your own timings will drift:

node scripts/0-validate.mjs --ad work/ad.json

It prints the beat map as a table and fails on a montage cut that is too long, a missing movement, a beat hold under 2.5s, a setup that does not accelerate, and a voiceover too long for the picture.

2. Show the user, and wait

Print the beat table, the joke, the pack line and the tagline. Ask for a yes. Nothing has been spent yet.

Never skip this.

3. Run the pipeline

node scripts/2-images.mjs --ad work/ad.json --dir work            # pack, sheets, shots, QC
node scripts/3-voice.mjs  --ad work/ad.json --dir work            # VO + the timeline
node scripts/4-motion.mjs --story work/ad.json --dir work --count 3
node scripts/5-sound.mjs  --story work/ad.json --dir work         # effects + mix
node scripts/6-render.mjs --ad work/ad.json --dir work --out work/ad.mp4

Steps 4 and 5 are the same scripts the animated-short skill uses, and they take --story because they never learn this is an ad: step 3 stamps each shot with the span it occupies, and that is the only thing they read.

To redo one shot:

node scripts/2-images.mjs --only 7 --fix "the sponge is soaked and foaming" --dir work

Text. Every frame is held to the no-letters rule except the pack, where the brand name is required and checked by eye against the spelling you gave. This is the one place an image model is allowed to write.

Motion. Two modes.

node scripts/4-motion.mjs --story work/ad.json --dir work --count 3   # the big beats
node scripts/4-motion.mjs --story work/ad.json --dir work --full      # everything that can move

--full animates every shot of 0.6s or more that has a motionPrompt. Stop-motion is supposed to move, and a still clay frame reads as a photograph of a puppet rather than a puppet, so for this format --full is usually worth the money. It costs about $3 more.

Two things it will not animate, both deliberate:

• Montage cuts. Three frames is below the threshold where anyone perceives

movement. The money is invisible.

• The pack. Image-to-video models erode printed text across a clip. Our

hero shot began as "Lifts mud in one wash" and ended as "Lifte nod lo ore wash" — on the one frame the whole ad exists for. Broadcast locks off the beauty shot for exactly this reason. --motionPack overrides it if you want to try; check the last frame, not the first.

If you swap the motion backend, note that the look-lock and the negative prompt are chosen per shot: a macro insert is told to stay photographic and is NOT given photorealistic as a negative, and a pack shot is allowed lettering while being told not to warp or misspell it. A single global negative list fights the format in both directions.

4. Hand it over

Give them work/ad.mp4. Never auto-post it anywhere.

What it costs

A 26s spot with 3 animated shots:

pack + sheets + ~25 shots + repaints   ~$1.20
voiceover                              ~$0.02
3 motion clips                         ~$1.26
  (or --full, 5 clips, ~$2.94)
ambience + effects                     ~$0.10
                                       ───────
                                        ~$2.60

All stills is about $1.30. Say the number before you spend it.

What fails

• Even cut lengths. The single most common failure. It stops being an ad.
• A montage that tells story. It is texture. Story goes in the beat shot.
• Animating the montage. Four times the cost, invisible.
• A pack described vaguely. Then it changes between the hero and the end

card, and the one frame the ad exists for is the one that looks fake.

• Letters anywhere but the pack. Invented lettering is the clearest tell

there is.

• Animating the pack. The brand name degrades over the clip. Check the LAST

frame of any shot you animate, never the first.

• Talking over the end card. Let the tagline land in silence.

Styles

looks.clay and looks.macro are plain strings — swap either and the pipeline is unchanged. The animated-short skill ships twelve tested look strings in its styles/ folder, each with a rendered proof image; the felt, crochet and papercraft ones all work here in place of clay.

One rule before you write your own: a physical medium needs the texture named, a counted palette, and a statement of what the medium is not. A photographic look needs none of that — lens and lighting terms carry it.