Website to claymation ad

The person gives you a company's website (or one product or community page). You deliver a ~20 s ad in a warm, stylized 3D claymation look, at 16:9, 1:1 and 9:16: five shots of a toy-like world, a narrator, a music bed, and kinetic titles and motion graphics in the company's own brand, ending on its own CTA button. One URL in, three files out, for about $5 of model spend.

The shape, agreed before anything is spent: 20 s, five shots of about 4 s, one narrator, about 48 spoken words. No real people: a recurring toy-like hero carries the story, and metaphors (a workshop, glowing blocks that snap into a chain, a tiny robot assistant) stand in for the product. Rounds and deliverables follow the core skill.

Steps

1. Capture the brand (guide-capture-a-brand, route A): the font family and weights, text and background colours, the accent, the CTA button's style and exact words, and the logo into refs/brand/. Keep the site's own copy beside it in source/: every fact in the ad comes from there.
2. Script and shot list. Five beats, one per shot: pain (a line with a number in it, such as "You've bookmarked fifty recipes. You still order takeout.") → the place (the brand name, "is where that changes") → what you get (two or three concrete things from the site) → proof (one number from the site) → CTA (the site's own CTA words). No invented reviews, numbers or claims. Gate: the script, the five-line shot list and the job folder, before any spend.
3. The look: shot 1 first (nano-banana-prompt-format). One style line, identical in every prompt: "warm, stylized 3D miniature, like a handmade toy-model world photographed with a tilt-shift macro lens; soft matte clay and painted wood; rounded shapes; shallow depth of field; toy-like figures with simple friendly faces. Not photoreal." Shot 1 on nano banana 2.1, 16:9, 2K, two takes. Gate: the pick locks the look and the hero.
4. The other four stills, all at once, on nano banana 2.1 edit with the locked still as the first image: "use it ONLY for the art style and the character… Do not copy its room." Keep the action in the middle third of every frame (it has to survive the 1:1 and 9:16), and end every prompt "No text, no logos." A prop that recurs (the robot) gets its earlier still as a reference, "use it ONLY for the robot", in every later shot. Gate: one review page, a pick per shot.
5. Vertical: redraw, don't crop. Redraw each picked still as a native 9:16 (nano-banana-restage-for-new-aspect), "as if the same camera had shot it in portrait… keep the same character, props, light, materials". Gate: a review page of the five. The 1:1 is a centre crop of the 16:9, shifted per shot where the subject sits off-centre; check a frame per shot.
6. Animate one, then the rest (h3-max-image-to-video, prompts per h3-prompt-format with "3D CG claymation" as the style): 768P, 4 s, a fixed seed, mouths closed and no speaking verbs. Quote the price, render one test shot, Gate: a yes on its motion, then the other nine (five 16:9, five 9:16) at once.
7. Voice and music (guide-cast-a-voice, elevenlabs-tts-v4, elevenlabs-generate-music), started together. Three library voices verified for eleven_v4, auditioned on the whole script with the with-timestamps TTS, so the picked audition is the take and its alignment gives the word times; pick for "upbeat, conversational, not salesy". One bed from a composition plan whose sections match the shots: sparse under the pain line, a swell as the doors open, full in the middle, one bright hit to end, no vocals. Gate: the voice and the bed on one page.
8. The cut (ffmpeg, free), one filter graph per aspect: each clip trimmed to its window (stretch one a little with setpts if short), cuts on the start of each VO line from the alignment (drop the [tag] characters), the music ducked under the voice with sidechaincompress, the clips' own sound at about 0.18, then loudness to −16 LUFS (ffmpeg-normalize-loudness). Check: −16 ±1 LUFS, and every word in a whisper-cli transcript of the mix (it needs a WAV, not the MP4). This untitled cut is rev1.
9. Titles and motion graphics (guide-add-titles-and-motion-graphics), one compositor spec per aspect over the rev1 cut, built by a Python script with the helpers in hn4a-compositor-in-scene-titles, cued from the word times. The look that was approved, in order:
   1. Hook title: the pain line as a big kinetic stack over a full-frame darkening. The words rise one by one, the number slams in huge in the accent colour with overshoot and a soft glow, then the last words land.
   2. Second line, same scene: as the stack clears, the next line builds word by word low in frame, smaller, the shade easing to a bottom-only gradient, its last word in the accent.
   3. That style for every remaining line, one accent word each. Write the phrases by hand on the word times, with a | where the narrow frames break the line.
   4. Logo lockup (mark plus wordmark in the brand font) at the top as the brand name is spoken, with no title for that line.
   5. Benefit chips: glass pills (white at 10%, a 22% hairline, blur) building on their words, one with an accent dot.
   6. Proof stat: the number in the brand's semibold with a short kicker under it, no title for that line.
   7. End card: a dark fade from the bottom, mark plus wordmark, the site's CTA button recreated, the URL small and dimmed, all in the lower third so the last shot's action stays visible. Let the last line's title run until the CTA word. Stills at every beat in all three sizes, then Gate: a review page with the three specs as live {"compose": …} candidates. Render only after the yes, and check the rendered frames at the logo and the end card.
10. Deliver per the core skill: the untitled cut and the titled one, three files each, in the round's folder.

Choices this workflow makes

• Claymation, stylized, never photoreal, and no real people: a toy hero and metaphors carry the product, so nothing on screen claims to be the real team or app.
• Every fact from the site; the CTA in the site's own words and button.
• Native 9:16 redraws. A 9:16 crop of a 16:9 frame is 608 px wide and looks soft full screen; a redraw is about $0.07 a shot.
• 768P for clay. Half the price of 1080P, and soft matte clay holds up scaled.
• Designed titles, not captions. One accent word per line, phrases written by hand.
• Auditions on the whole script with timings, so the pick needs no second read and no paid transcript.

Dead ends

• Karaoke captions: tried and dropped. The titles read as designed, the captions as subtitles.
• Automatic phrase grouping split a line on every small pause. Write phrases by hand.
• Sizes given by objects: "a robot the size of a teacup" came back as a literal teacup. Give sizes by body parts ("as tall as her knee").
• A laptop drew a real manufacturer's logo despite the prompt; pick the take without it rather than editing it out.
• A dropped connection on a parallel still: the file had usually landed. Look in the folder before re-running.
• The VO overran the cut into the end card; ending the last title early left a gap. Run it until the CTA word.

Worked example

Globex, a recipe-planning web app, from its homepage alone. 16:9 1920×1080, 1:1 1080×1080 and 9:16 1080×1920, 20 s, 24 fps. Brand: Brand Sans 400/600, a near-black background, one warm orange accent on links, a pill CTA "Start planning". One hero (a small clay cook) and a recurring tiny robot helper.

Each still redrawn as native 9:16; the 1:1 crop shifted on shots 3 and 5. Ten 4 s clips on H3 Max 768P after one test. Three voices auditioned on the whole script, 18.5–20 s each; the picked take used unedited. Note on the first titled pass: the karaoke captions read as subtitles, so they were replaced by designed titles.

Spend: stills about $1.25 (including the redraws), animation about $3.60, voice and music about $0.20: about $5. The cut, titles and renders were free.