Website to claymation ad

Give it a company's website (or a community/product page) and get a ~20 s ad in a warm stylized 3D claymation look, delivered at 16:9, 1:1 and 9:16, with kinetic titles and motion graphics in the company's own brand: their font, colours, logo and CTA button. One brief in, three files out, for about $5 of model spend. Made for outreach gifts, agency demos and "my agent made this" posts.

Shape (agree it first)

20 s, 5 shots of ~4 s, one narrator, a music bed. ~48 spoken words.
Script in five beats, one per shot: pain ("You've watched a hundred AI videos. You still haven't built anything.") → the place (the brand name, "is where that changes") → what you get (2–3 concrete things from the site) → proof (one number from the site) → CTA (the site's own CTA words).
Every fact comes from the site. No invented reviews, numbers or claims.
No real people. A recurring toy-like hero carries the story; metaphors (a workshop, glowing blocks that snap into a chain, a tiny robot assistant) stand in for the product.
Show the person the script and 5-line shot list, and agree the job folder, before spending anything.

1. Capture the brand (free)

Open the site in the browser and run in the page:

const cs=e=>{const s=getComputedStyle(e);return{font:s.fontFamily,weight:s.fontWeight,color:s.color,bg:s.backgroundColor}};
const colors={};[...document.querySelectorAll('*')].slice(0,3000).forEach(e=>{const s=getComputedStyle(e);[s.color,s.backgroundColor].forEach(c=>{if(c!=='rgba(0, 0, 0, 0)')colors[c]=(colors[c]||0)+1})});
({body:cs(document.body),h1:[...document.querySelectorAll('h1,h2')].slice(0,3).map(cs),
 buttons:[...document.querySelectorAll('a[class*=btn],button')].slice(0,4).map(e=>({text:e.innerText,...cs(e)})),
 colors:Object.entries(colors).sort((a,b)=>b[1]-a[1]).slice(0,15),
 fonts:[...document.fonts].map(f=>f.family+' '+f.weight),
 icons:[...document.querySelectorAll('link[rel*=icon],meta[property="og:image"]')].map(l=>l.href||l.content)})

Write down: the font family and weights, text and background colours, the one accent colour (often rare in the count but on links/plus signs), the CTA button's style and exact words, and the logo (the 180–192 px apple-touch/favicon PNG is the cleanest mark). Download the logo into refs/brand/; load the font from Google Fonts in the spec.

2. Stills (nano banana 2.1, ~$0.85)

Load 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 first, fal/google/nano-banana-2.1, 16:9, 2K, 2 takes. Show it; the pick locks the look and the hero. Then shots 2–5 on fal/google/nano-banana-2.1/edit with the locked still as image 1: "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. End every prompt "No text, no logos." (a laptop came back with an Apple logo; pick the take without it).
Give sizes by body parts, never objects: "a robot the size of a teacup" came back as a literal teacup. A prop that recurs (the robot) gets its earlier still passed as a reference image with "use it ONLY for the robot", in every later shot.
Run shots in parallel; if a call drops its connection, look in the folder before re-running: the file has usually landed.
One review page for all five; the person picks per shot.

3. Vertical: redraw, don't crop

A 9:16 crop of a 16:9 frame is 608 px wide and looks soft full-screen. Redraw each picked still as native 9:16 (nano-banana-restage-for-new-aspect, one 2.1 edit each, ~$0.07): "as if the same camera had shot it in portrait… keep the same character, props, light, materials". Review them. The 1:1 is a centre crop of 16:9; shift the crop per shot where the subject sits off-centre (check a frame per shot).

4. Animate (H3 Max, ~$3.60 for ten clips)

fal/minimax/h3-max/image-to-video, 768P, duration: 4, prompt_expansion_mode: "disabled", seed fixed, prompts in h3-prompt-format's image-to-video shape ("3D CG claymation" as the style). 768P is $0.088/s; 1080P is $0.176/s (double), and 768P holds up for soft clay. Quote the price, render one test shot, get a yes, then the other nine (5 × 16:9 + 5 × 9:16) with wait: false and tasks(wait: true).

Clips come back ~4.5 s. Mouths: "Lips stay closed" and no speaking verbs.
Saturated wardrobe drifts brighter over a clip; it reads as the character stepping into warmer light, so leave it unless it jumps.

5. Voice and music (~$0.20)

Three library voices verified for eleven_v4, auditioned on the whole script with elevenlabs/text-to-speech/with-timestamps (it returns a character alignment JSON you need for the titles). Pick for "upbeat, conversational, not salesy". All three should land 18.5–20 s; keep the take the person picks, unedited.
Music: elevenlabs/music with a composition_plan whose sections match the shots (sparse intro under the pain line, a swell as the doors open, full in the middle, one bright hit to end), "vocals" in the negative styles, respect_sections_durations.

6. The cut (ffmpeg, free)

Read the word times from the alignment JSON (drop the [tag] characters) and put each cut on the start of a VO line. One filter graph per aspect in a file, run with -/filter_complex <file> (ffmpeg 8 dropped -filter_complex_script): trim each clip to its window (setpts to stretch one a little), scale, concat; VO + music (sidechaincompress keyed on the VO) + the clips' own sound at ~0.18; loudnorm I=-16 TP=-1.5. 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; titles go on top of it.

7. Titles and motion graphics (hn4a-compositor, free)

One spec per aspect over the rev1 cut, built by a Python script (hn4a-compositor-in-scene-titles, titles_fx.py), the mix as the spec's audio. The look that landed with the client, in order:

Hook title. The pain line as a big kinetic stack over a full-frame darkening: "You've watched" rises word by word, the number slams in huge in the accent colour (pop_keys, over=1.45, with a soft glow), then "AI videos." lands.
Second beat, 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, the last word in the accent colour.
That style for every remaining line, not bouncing karaoke captions (tried and dropped: the titles read as designed, the captions as subtitles). One accent word per line. Write the phrases by hand on the word times with a | where narrow frames break the line; automatic phrase grouping splits on every small pause.
Logo lockup (mark + wordmark in the brand font) at the top as the brand name is spoken; no title for that line.
Benefit chips: glass pills (white at 10%, a 22% hairline, blur) building on the words, one with an accent dot.
Proof stat: the number in the brand's semibold, a short kicker under it; no title for that line either.
End card: a dark fade from the bottom, mark + wordmark, the site's own CTA button recreated (same shape, colours and words), the URL small and dimmed. Keep it in the lower third so the last shot's action stays visible; let the last line's title run until the CTA word, since the VO often overruns the cut to the end card.

Traps: an <img src> path inside a text layer's html loads in the preview but not in render; inline the logo as a data: URI. Name the font family on every text layer. Stills at every beat in all three sizes, then a review page with the three specs as {"compose": …} candidates playing live; render only after the yes, and check the rendered frames (logo, end card) before delivery.

Deliver

rev1/ the untitled cut, rev2/ the titled one, three files each, on an approve review page. Typical spend: stills ~$1.25, animation ~$3.60, voice + music ~$0.20 → ~$5.

Hypernatural

recipe-website-to-claymation-ad