Launch films
A film here is one recorded session of the real product, cut into a handful of scenes. Each scene is a flat colour field, the product large on it, and one heavy headline of two to five words with one phrase boxed. No stock photos, no actors, no 3D. The model is the release clips posted by accounts such as @HermesWatcher: claim, typed request, watch it work, result, handle.
The kit root is two levels above this file ($KIT; in a consuming repo usually
trailers/kit).
$KIT/scripts/new-trailer.sh product|feature <dir> ["Brand"] [url] # scaffold: trailer.json, record.mjs, storyboard.md
node $KIT/scripts/record.mjs <dir> [--only clip] # play record.mjs in the real app -> public/*.mp4 + clips.json
node $KIT/scripts/check-spec.mjs <dir> # validate + timeline + [craft] warnings
$KIT/scripts/render.sh <dir> [landscape|square|vertical|all] [--gif] [--poster] [--draft] [--strict]
node $KIT/scripts/review.mjs <dir> # contact sheets of the render: LOOK at them
$KIT/scripts/studio.sh <dir> # live preview :3000
node $KIT/scripts/grid.mjs <frame.png> /tmp/g.png [20] # read coordinates off a frame
node $KIT/scripts/beats.mjs <dir>/public/music.mp3 # only if the film has music
What went wrong before (don't repeat it)
The owner rejected two generations of trailers from this kit. Their words on the second: "Everything is so incoherent, a man suddenly turns into a woman, a woman on a smartphone operating a desktop UI, no story, so chaotic and AI slop in general." Both generations passed validation. The causes:
- Stock photos. A different stranger in a different room in every scene, with a small product card floating beside them. Nothing connected one scene to the next, and the device in the photo didn't match the UI shown.
- Fragments, not a job. Three-second crops from unrelated recordings with unrelated data. The viewer never saw a task start and finish.
- A feature tour. Each scene announced another feature; no scene followed from the one before.
- Unreadable UI, equal-length scenes, a crossfade on every cut.
The rules
- One session, one person, one job. Decide who is using the product and what they get done in the next two minutes. Record that, start to finish, in one browser session, so the task typed in scene 2 is the task ticked in scene 4 and the one shown under its goal in scene 5. Continuity of data is the story. If a scene could be removed without breaking the chain, it's a tour stop: cut it.
- One promise, told in the product's own method. Write the sentence the
viewer should believe afterwards as
promise. Scene 1 states it; the end card repeats it. If the product has a named method or steps (Life Console: Strategise → Distribute → Execute), that method is the film's structure: one scene per step, in order, each labelled with akicker("1 · Strategise"), and the opener lists the steps. Ask the owner what the method is and which screen they are proudest of; give that screen the most room. - The product is the picture. Every scene except the end card shows the real UI, large, doing something: text typed, a slot filled, a timer started. No stock photos, no illustrations of people, no device mock-ups.
- The data on screen must agree with the words. Read every large value
in the shot against the headline. The owner rejected a dashboard reading
"Free until 18:30" under the line "tells you what to do now". Set the
clock, seed the data and tick off earlier tasks in
record.mjsso that the state on screen is the state the headline describes: a task running, not a gap; nothing overdue beside a task marked done; no deadline shouting "due today" for something already finished. - Frame a whole component. The camera (
view) shows one complete panel: never a window that cuts words in half at its edge. The thing the headline claims must be readable at phone size. Readphone.pngto check. - Few words, heavy, one boxed. Two to five words, broken into short lines
(about ten characters a line keeps the type big). Box the word that carries
the claim:
It finds [the time].The headline says what it means for the viewer; the picture shows it happening. Never caption the picture. - Show the request. When the person types or asks something, show it as a
promptbubble typed over the product, then cut to the product's answer. - Colour is the edit. Paper for the opening and the end; one or two brand
colour fields in between; the product full-frame (
layout: "full") at least once. Hard cuts only. Alternate which side the words sit on. Vary scene lengths; the longest goes to the scene where the most happens. - The logo opens and closes it, large. Ask the owner for the logo file;
don't assume the app icon is it. Open with the icon and name
(
brandMark, frombrand.logo); close with the full logo artwork, large, on theendscene (brand.lockup). The owner chose that order: icon first, full logo as the pay-off. An icon alone does not count as using the logo. - End on a hero shot.
endwithmedia: the logo, the promise, one button and the URL, with the product rising large behind them. Use the product's best state for it (the same rule about data applies). - Sound is the owner's call. Ask once. Silent is acceptable (set
"silent": true); then the words carry everything. With music, pick the track first and cut to it (beats.mjs).
check-spec.mjs prints [craft] warnings for what it can measure. Fix each or
be able to say in one sentence why it doesn't apply (an establishing shot of the
whole app may have small labels; the big type still has to read). Only the
review step can tell you the film is good.
Workflow
- Brief. Read the repo first (README, landing page, changelog, seed or demo data). Then ask the owner what you can't find: who it's for, the one promise, where it will be posted, sound or silent. They have said "you can ask me all the questions": a wrong guess costs a full cut.
- Storyboard. Fill
<dir>/storyboard.md: the viewer, the promise, then the session as a chain: they do X → the product does Y → which lets them do Z. Five scenes or fewer plus the end card. Use the product's own demo data if it has a coherent persona; otherwise write one and seed it. - Record. Run the app locally, write
<dir>/record.mjs(references/recording.md), runrecord.mjs. Look at a strip of every clip before going on (ffmpeg -i clip.mp4 -vf "fps=1,scale=640:-1,tile=5x3" -frames:v 1 strip.png): stray dialogs, toasts and open drawers show up here. - Spec. Write
trailer.json(references/spec.md). Onedemoscene per clip; takestartFrom, step times and zoom times from the marks inclips.json. Read camera centres off a frame withgrid.mjs; never guess them. - Check.
check-spec.mjsuntil no errors. - Draft and review. Not optional.
render.sh <dir> --draft, thenreview.mjs <dir>, then opensheet.pngandphone.pngand answerreferences/review.md. Expect two or three rounds: reframe, shorten lines, re-record a clip, cut a scene. - Final.
render.sh <dir> all --gif --poster --strict. Review the vertical cut separately. Report the paths, what you cut and why, and what is still weak.
The shape
Lengths are examples. A product film runs 20–35s; a feature clip 12–25s.
| # | scene | example | length |
|---|---|---|---|
| 1 | demo, paper, tilt, brandMark, steps | Logo + "Bring systems into [your life]." + 1 Strategise 2 Distribute 3 Execute, beside the live app | 4s |
| 2 | demo, colour field, layout: "top", kicker | "1 · Strategise" / "Start from [the goal]." over the goal flow lighting up | 6s |
| 3 | demo, second colour, kicker, steps | "2 · Distribute" / "Level [the week]." while an overloaded day is evened out and time-boxed | 6–7s |
| 4 | demo, layout: "full", kicker | "3 · Execute" / "Do the one thing [now]." over the running task | 4–5s |
| 5 | end with media | hero shot: logo, the promise, Try it now, URL | 4–5s |
A product without a named method uses the same shape with a typed request in
the middle (prompt: "Just type it."), as in life-console/trailers/launch-v2.
Feature clip (existing users). Skip the opener's explanation: scene 1 is "[Product] can now [do X]." beside the feature, then the request, the result, the end card. Three scenes plus the end card is enough.
How-to (60–80s, only on request). Numbered steps: put 1/5 Open Settings
style labels in the headline, one demo scene per step, a recap with steps.
Worked example: life-console/trailers/launch-v3 (trailer.json, record.mjs,
seed.mjs, storyboard.md).
Older scene types
moment (stock photo + floating widget), widget, screen, the 3D scenes,
features, montage and the rest still render, and are documented in
references/spec.md and references/moments.md. Don't use them unless the
owner asks for that look by name; never mix them with demo scenes.
Extending the kit
- New look:
theme(paper and ink colours, accent, font). - Something
democan't do: add a prop toframework/src/scenes/Demo.tsx, its type inspec.ts, its checks inscripts/check-spec.mjs, and document it inreferences/spec.md. Extend the kit; don't fork it per film. - A rule learned the hard way: a
[craft]warning incheck-spec.mjsif it can be measured, a question inreferences/review.mdif it can only be seen.
Then run npm test in $KIT.