Question film
Explicit user instructions take priority over this workflow. Work only on authorized products and sample data. Do not request passwords, API keys, payment data, government identifiers, health records or other sensitive data. Explain missing local dependencies before installing them in the working project. Network access is for the authorized product and its licensed assets, plus the declared npm dependency; respect access controls. Render one job at a time and wait for completion. Keep publishing, sending and deployment outside this task.
If several skills could do this: when the request doesn't name this skill and your agent also has other omnigami skills that make the same kind of thing (other launch films, shots, decks or mockups), list them by name and ask which one the user wants before you start. If it's named, or it's the only one installed, go ahead.
For Japanese, Korean or Chinese product pages, keep the copy in that language and set the film’s lang to the page’s <html lang>; bundled Noto faces and phrase breaks handle the text.
A 30-second tall film for a product you ask things of. The whole frame is the product's input field. The user's question fills it, big and flush left, with the field's send button in its corner. Sent, the question folds up into a headline band at the top and the product's real answer rises under it, so every answer is shown with the question it answers. The next question drops down over the answer as a new band. A real button in an answer ("Book it") grows into the next request. The last band closes back into the product's own input, its line is typed in, and the film ends on the lockup. Flat color fields, one heavy sans, real screens. 30 fps on a 132 BPM grid, tall (1080 × 1920) only.
You work by adapting a finished film. reference/ is the complete film for Ostlery (a fictional stay-booking
assistant made for omnigami's examples; its story is in reference/STORY.md). Keep its machinery: the easing, the
timeline, the type rules, the fold, the reopen, the button opening, the close and the film's own checks. Its nouns
become the user's: the input, the questions, the answers, the button, the colors, the fonts, the line, the mark and
the wordmark. Every setting you change is labeled ADAPT in reference/index.html; everything labeled KEEP stays.
Read style.md before you plan and critique.md before you judge anything.
What it needs
- The product itself, running: its app, a demo or its source you can serve. Every screen in the film is a capture of the real product with sample data, never a screen drawn for the film.
- Its input field: the field the user types a request into, with its own send button (a composer, a search box with a go button, a prompt box). The film opens on the page that holds it and closes back into it.
- Three answers it really gives, each a part of a real screen: a first answer to a long request (a result list, a search with its matches), a second that picks or explains (one result with the product's reply), and a third that follows a real button in the second answer (a confirmation, a booked or sent state). A fourth answer, to a short practical question, ends the story (directions, a detail, a next step).
- The words: the hook typed into the input (the product's own prompt, about four words, its name in it), three requests a real user would type, the button's own label, and the product's line (three to six words).
- Its brand: the mark as an SVG, the wordmark (its face, weight and tracking, or an outlined SVG), the interface face (files with their licenses), the brand color, the ink, the ground, a quiet fill and a success color.
Read the site and the app before asking; then ask once, in one message, for what they can't answer. If the user wants no questions or an unattended run, decide from what you can reach and say what you decided.
Set up
Check Node 20+, ffmpeg and a Chrome-family browser (scripts/chrome.mjs finds it; CHROME_PATH overrides). Explain
anything missing and stop. Work in question-film/ in the current directory; set SKILL_DIR to this file's folder
first:
SKILL_DIR=/path/to/question-film # the folder this SKILL.md is in
mkdir question-film && cd question-film
cp -R "$SKILL_DIR/reference" reference && cp -R "$SKILL_DIR/reference" film && cp -R "$SKILL_DIR/scripts" scripts && cp -R "$SKILL_DIR/music" music
rm -rf reference/out film/out film/STORY.md
npm init -y >/dev/null && npm i -D 'puppeteer-core@^24'
node scripts/render.mjs reference --size 9:16 --times key --sheet --out out/reference
The film is tall only: pass --size 9:16 to every render.mjs command (it defaults to 16:9). reference/ stays
untouched, to compare against. You edit film/: its index.html and its assets/ (the product's captures in
assets/screens/, its photographs or media in assets/stays/ if its answers show any, its mark, its fonts beside
their licenses; delete the reference's once yours load). Your STORY.md and PRODUCT.md go at the top of
question-film/. Look at out/reference/9x16-sheet.png now: that is the bar.
1. Read the product
node scripts/capture.mjs <url> captureand read every screenshot andcapture/brand.json. Then open the product itself and use it: type a request into its input, follow the answer, press the button the answer offers. A local app: serve it (python3 -m http.server) and use its address.- Write
PRODUCT.md: what it does in one sentence, its voice, its fonts and licenses, its line, its input field (fill, corner radius, hairline, placeholder, send button: color, shape, icon, its idle and ready states), the answers you found and the button in the second one, and its colors by role (the film's tokens,C):paperthe app's ground,cardthe input's fill,inkits text,softits placeholder gray,linethe input's hairline;emberthe brand color andember2that color pressed (the names are the reference's; keep the role names);sanda quiet fill the app uses (an idle button, a muted panel),rulea light pressed shade;okits success color (the field the button opens into: the button's result). Each with where you found it, and the date.
2. The story and the words
The story is four requests and their answers, in the order someone really uses the product:
| Request | Field | What it is |
|---|---|---|
The hook (HOOK_TEXT, HOOK.text) | the ground, ink type, the name in the brand color | the product's own prompt, typed into its input: "Tell Ostlery the trip." |
1 (q1) | the brand color, ground-colored type | the long first request, the way a user would type it: "A weekend lake cabin, with our dog, on a budget" |
2 (q2) | ink, ground-colored type | a short follow-up that asks it to choose or explain: "Which one's best?" |
3 (Q3_TEXT) | the success color | the real button in answer 2, by its own label: "Book it" |
4 (q4) | the quiet fill, ink type, one word in the brand color | a short practical question: "How do we get in?" |
The line (LINE_TEXT) | typed into the real input at the end | the product's line: "Plan a stay in a sentence." |
- Requests are the user's words, in the user's voice: plain, specific, the way someone types to the product. The answers are the product's real screens with its sample data; the film never sets a result, price or name from the sample data as a statement.
- Lengths (the film's checks fail beyond them): request 1 six to ten words (up to 64 characters), set in up to five lines; requests 2 and 4 two to four words (up to 32 characters), in two or three lines; the button's label one or two short words; the hook about four words (up to about 26 characters, typed at 0.045 s a letter before the box opens at 3.18 s); the line up to about 26 characters (typed at 0.04 s a letter, done before the press at 27.17 s). Each request also folds into a headline of at most two lines at 85–120 px.
- Break the lines by hand with
/("A weekend / lake cabin, / with our / dog, on a / budget"), so each line is a phrase and the fold keeps reading order: a word may only travel to its own line or a line above, never past a word that stays. The film's check ("no word passes through another") fails a break that makes words cross in the fold; re-break the request, never the check. - One word of request 4 may be marked
*like this*to set in the brand color (the reference marks "in?"). The hook's key word (the product's name) is marked the same way inHOOK.text.
Write STORY.md in the shape of reference/STORY.md: the TL;DR, what each beat says, the beat sheet with your input,
answers, button and fields, and every word with its source. Show the user the requests, the answers you'll capture and
the button in a few lines and ask for one review, unless they asked for an unattended run. Never take silence as
approval.
3. Capture the real screens
Write a plan for scripts/screens.mjs (the format is at the top of the script) at a 390 × 844 phone viewport, and run
it into film/assets/screens/ (node scripts/screens.mjs plan.json film/assets/screens). The film draws captures by
name from screens.json, so keep the reference's names:
| Name | What | Steps |
|---|---|---|
home | the page that holds the input, whole viewport, scale 3 | capture with full |
composer | the empty input, scale 4 | capture around the input |
ask-01 … ask-NN | the hook typed a letter at a time | type into the input with capture: "ask", around the input |
line-01 … line-NN | the line typed the same way | type with capture: "line" |
search | answer 1's lead panel (the reference: the search card) | capture around it after sending request 1 |
a2 | answer 2: the chosen result, the reply and its buttons, as one panel | capture around the response |
a3 | answer 3: the state the button leads to (Booked, its card) | press the button, then capture |
trip-info | answer 4's panel (the reference: the trip page's Drive and Getting in rows) | capture around it |
Record the input's rect (rects) so COMP, TXT and SEND match it: COMP is the input's box in CSS px (x, y, w,
h, its radius), TXT its text's left inset, size and first baseline, SEND the send button's center and radius
inside it. Before capturing, check its labels sit centered: node scripts/centring.mjs plan.json, and report what it
lists. Answers that show photographs (the reference's six stays, the cabin and its room) use the product's own images
in assets/stays/ (STAYS, FOCUS); an answer without photographs rises as captured panels only.
4. Adapt film/index.html
- Tokens and faces (
C,SANS,SERIF): the brand's colors by role (step 1).SANSis the interface face, set at 880 for every request and headline: a variable face with a heavy end (the reference's Figtree runs 300–900); a face without one: its heaviest weight, and say so.SERIFis the wordmark's face. Load both inprepare(). - The mark and icons (
MARK,ICON_UP,ICON_CHECK): the SVG's path data copied exactly (scaled into the 100-unit box if its viewBox differs); the app's own send and check icons, copied exactly. - The requests (
HOOK,REQ,defaults(),fields,Q3_TEXT): the words from step 2; each request's field, type, accent and send-button colors (bg,ink,acc,btn,arrow,darkfor its press);bh, each band's height (the first, two lines, 440; the rest 330). - The answers (
answer1…answer4,A1…A4,A2_ROWS,A2_PILL,A3,PILL): the product's captures, at 2.6–2.7× their CSS size (a 390 px phone page at the frame's width is 2.77×), centered under the band. Keep how they arrive (KEEP): every part rises from below into the space the band leaves, staggered (photos 0.06 s apart, left to right and row by row; a response's rows 0.045 s apart), onEXPO_OUT, never a fade or a slide from the side (answer 2's carousel slides in from the right inside its panel, because the product's carousel scrolls that way). Measure the button in thea2capture intoA2_PILL(its CSS rect) and redraw it the app's way indrawPill, so it can be pressed and grown. - Answer 3's state change (
bookPill,T.booked,T.done): the product's own in-progress state, then its done state, one label at a time, then the captured done panel. If the product shows no progress state, cut that step and rise the done panel onT.booked. - The lockup (
LOCK,endLock,sceneEnd,prepare'sMW): keep the proportions (the wordmark's size 124:98.6 of the mark's drawn height, the gap 26:98.6, the baseline on the mark's foot, the mark a touch above the middle: the film checks them). The word, its face and weight are the brand's; its tracking is the brand's own (the reference sets Ostlery's −0.0174 em, as its other films do: a wordmark set looser than the brand sets it was sent back). An outlined wordmark: draw its path instead offillText. - KEEP: the easing curves,
T(change a time only with its neighbors,style.md),MARGIN,TRACK,LEAD,WT, the layout and its breaks,foldK,flipWord,riseWord, the fold, the reopen, the button opening, the iris, the close, the cues andselfChecks.
The rules the founder set for the fold (all KEEP; the film's checks hold the first four)
- The words slide across first and rise only as the band comes up.
foldK: across (with the size) from the press over 70% of the move; the rise waits 0.3 s, until the band has come up close under the words, then lands with it. - One shared scale. Every word of a request on one progress, one curve, one start: no per-word or per-line stagger in the fold, so all words are the same size on every frame.
- No word passes through another, in any fold, in the hook's row-to-column move or in the button's opening.
- Enough leading that descenders clear the line below (
LEAD = 0.98; the check measures the ink). - Answers rise from below, staggered, into the space the band leaves; never cross-fade.
- The first request rises inside the swelling send-button disc, so the opening disc pulls the words in (they start 0.2 s into the iris).
- The composer close shows no placeholder: the closing band becomes the empty input with only its idle send button; the line is then typed into it.
- The wordmark's tracking is the brand's.
5. Check, critique, revise
node scripts/render.mjs film --size 9:16 --check: frames match when rendered twice in different orders, and the film's own checks pass (it prints each failure with its numbers, and the count when all pass): the lockup's proportions and padding, every request and headline inside its margins and band, every word inside the frame on every fold frame, no word passing through another, the leading, and the first request starting inside the opening disc. Fix every failure in the film, never in the check; if you replace a step, change its check to match.node scripts/render.mjs film --size 9:16 --stress(the words at their longest and shortest),--times key --sheet --out out/round<N>,--phone, and--review(strips around every cut and key time). Lay your sheet beside the reference's (node scripts/compare.mjs reference film out/round<N>/compare.png) and look at the images.- Score every still and strip on
critique.md. Get a fresh judgement if your agent can start a sub-agent without this session's memory: give it onlycritique.md,style.mdand the round's images beside the reference's; merge what it finds. Otherwise rereadcritique.mdand score from the images alone, as a stranger would. - Fix, and keep notes in
out/round<N>/NOTES.md. Stop when every line scores 8 or more; after three rounds, show the user what remains.
6. Render
node scripts/render.mjs film --size 9:16 --video --sound --audio music/song.mp3 --stems music/song-stems --bpm 132 --land 27.27 --poster 29 --out out
This writes out/9x16.mp4 (30 fps, no motion blur, as the reference), its poster on the lockup, and the soundtrack:
the skill's song (music/song.mp3, 132 BPM, the film's grid) from its first downbeat, its last hit landing on
--land (T.endTurn, 27.27 s: the composer turns away and the mark turns in), where its drums stop and the rest
closes and fades (from its stems), with the kit's quiet effects on the film's cues under it (ticks on keys, taps on
presses, soft swishes on the big moves, one pop on the done state). If you move T.endTurn, move --land to match.
--music instead of --audio … --land … gives the synthesized score; the user's own song works with
--audio their-song.wav --land 27.27 (without --bpm 132). Run it in the foreground and wait for it. Then
node scripts/motion.mjs out/9x16.mp4 <from> <to> over each fold (press to press + 1.8 s), each reopen, the button
opening and the close; critique.md lists the flags the machinery raises by design.
7. Deliver
Run node scripts/preview.mjs film and give the address it prints: it plays the film and edits the requests (a saved
edit needs a new render to reach the MP4). Write README.md (the files, how to preview, how to change a request, how
to render), out/CAPTION.md (a post under 280 characters in the product's words, with its URL) and out/REVIEW.md (the
scores, round by round). Hand over the MP4, the poster, the caption, the check results and the one change that would
most improve the film.
Hard rules
- Only the user's product. Its real input, its real answers captured from the running product, its real button, its mark exactly, its words, its colors, its fonts with licenses. No screen drawn for the film; sample data stays inside the screens.
- Nothing from the inspiring film: none of its footage, frames, words, logo, colors, music or app. Its grammar only.
seek(t, values)paints from its arguments only: no clocks, unseeded randomness, timers or network; CPU-backed canvases (willReadFrequently: true),ctx.reset()every frame. Every file the film loads lives insidefilm/.
Files
reference/: the Ostlery film (index.html), its story (STORY.md), its captures (assets/screens/, from Ostlery's working app), its stays' photographs (assets/stays/), its mark (assets/mark.svg), its fonts (Figtree and Newsreader, SIL Open Font License,assets/fonts/) andkit.js. To watch it with its song:node scripts/render.mjs reference --size 9:16 --video --sound --audio music/song.mp3 --stems music/song-stems --bpm 132 --land 27.27 --out out/reference. Ostlery is a fictional company made for omnigami's examples.style.md: the grammar, with numbers.critique.md: the review, every rule with how to check it.music/: the skill's song and its stems.scripts/:render.mjs(stills, strips, the quick test, checks, MP4s),compare.mjs,motion.mjs,screens.mjs(the product's real screens),capture.mjs(brand from a URL),centring.mjs,preview.mjs,music.mjs,song.mjs,sfx.mjsandchrome.mjs.
This skill is omnigami's, © 2026 Auberon LLC. Use what it makes for anything, commercial work included, and change the skill for your own use; don't redistribute or sell the skill. Full terms: LICENSE.md