Workflow: launch, promo or release film from a repo, a site or a brief

Docs / skills/showtime/references/workflows/launch-video.md

Read this when the user wants a launch, promo, announcement, "hero" or release video for a product ("make a launch video for this repo", "a 30-second promo for acme.dev I can post on X", "a video for v2.0"). For a what's-new tour of several changes in one release use changelog-video.md; for a click-by-click walkthrough, tutorial.md; for a teaser that withholds the product, trailer.md.

The bar: what premium measures like#

Measured on launch films that viewers rate premium, and on ours that read "choppy and cheap":

	premium	cheap
story	4-6 scenes in 20-45 s, one message each	12-17 layouts, a new grid every 3 s
scene changes	a match or a soft dissolve (at most one motivated camera move); 0-5 hard cuts	pushes, wipes and cuts that replace the whole frame
holds	each scene lands, then holds 2.5-5 s with only slow drift	content arrives and leaves mid-motion
type	one display face + one mono, headline >= 6 % of the frame height, info text >= 3.5 %	five styles at once, 1.4-2 % commands
colour	one ground for the film, one accent	a new palette per scene, saturated edges
motion	ease-out entrances, eased camera moves, nothing bounces	punch-in zooms, shaky or bouncy elements, pops
music	a real produced track with dynamics; the scene changes on its phrases, the name on its swell	a flat synthetic bed, an effect on every cut

The launch template, the camera transitions and showtime audio cuts build the left column by default. showtime qa measures it (rhythm line: layouts, how each change is made, hard cuts, still stretches, music dynamics) and warns on the right column.

Defaults (state them in the opening line, don't ask)#

30 s (20 s when the request says short, 45 s for a site hero), 16:9 at 1920x1080 30 fps (X and LinkedIn play it as is), plus a 9:16 cut of the same page only for a vertical-first platform (Reels, TikTok, Shorts) or when asked; template launch (5 scenes); a produced catalog track cut to the film (showtime audio cuts), no sound effects; no voice-over; no captions (there is no voice). Ask at most one question, only when nothing hints at platform or length.

Scene grammar#

#	scene	job	length	on screen	hands over by
1	hook	the promise or the problem	3.5-6 s	one line, >= 9 % of the frame height, the key word in the accent, complete at frame 0 (it is the poster)	`through`: the camera flies through a letter of the key word (`data-portal="counter"`) into scene 2
2	verb	the product doing its one job	6-8 s	the product window (terminal with a real command and its real output, or a real capture or demo clip) + one headline of 3-6 words	`match`: the window stays; its content and the headline change
3	proof	the second benefit	5-7 s	the same window, the next command or screen; one result	`match` again
4	proof	the third benefit	5-7 s	as 3	`blur-dissolve 0.9` (or `dip`)
5	end card	name, value, how to get it	4.5-7 s, from the music's swell	name (+ version), the one-line value in the product's words, the install command or CTA exactly as the README gives it, the URL or repo, the music credit line (filled in automatically)	holds to the end

20 s: 4 scenes (drop scene 4). 45 s: 6 scenes (add a "range" scene before the end card: several real results side by side, entered with a dissolve).
Each proof is "verb, then result": the command types, the output lands, the result line is marked, and it holds at least 1.5 s. The camera does not move. One result per scene, never a list of features.
A web product: the window is a browser-frame with a real capture (data-src), a mobile app a device-frame, a video tool its real output clips (VP9 proxies, stage-api.md). Keep data-match="win" on the window in every scene that shows it.
Scene 1 holds a line only the product can say; "Introducing X" is a label, not a hook (story.md §3).

Camera and motion rules#

Every camera move must have a reason; if you can't say it, cut it. Viewers read an unmotivated zoom or pan as noise ("the zooms didn't make sense"); a calm, still frame with lively content reads as premium.

Default handoffs: match (the product window carries across, only its content changes) and a soft blur-dissolve or dip (0.8-1 s) where the picture really changes (into the end card). At most one hard cut, on a downbeat.
At most ONE fly-through per film (through): from the hook into the product, and only when the hook's key word is big (>= 9 % of the frame height) and the product appears inside its letter. Otherwise a dissolve. Never a second one, never pan, push, whip-pan, zoom-through, glitch, flash or shader transitions in a launch film.
The camera is still inside every scene: no pushes, zooms or drift on the hook, the proof scenes or the end card. Motion comes from the content: the command typing, the output landing line by line, the result line being marked a beat later, the end card's words arriving. (The camera component stays available for a real reason, e.g. a UI detail too small to read: say the reason in the plan.)
Never: punch-in zooms, shake, bounce or overshoot (back, springs under 1), floating or breathing loops on text, any text leaving the frame.
Entrances ease out over 0.4-0.9 s; words arrive 0.1 s apart; nothing exits on its own (the transition is the exit).
Holds are fine: a settled result or the still hook for 2.5-5 s is how premium films breathe. Every text holds long enough to read twice (pacing.md); the end card holds 3 s after its last element lands.

Type and colour#

One display family and one mono: the product's own fonts (brand.json, the site's CSS), else Geist and Geist Mono (the template's). At most three sizes on screen.
Hook >= 9 % of the frame height at 16:9 (13 % of the width at 9:16, filling it on 2-3 lines), held still and complete; headlines >= 6.5 %, commands and UI text >= 3.5 % at 16:9. Terminals and command lines carry data-st="fit": their type shrinks until the longest line fits the window at every size (at 9:16 real commands land near 2 % of the height; keep commands short).
One message at a time: the headline and the window that proves it. No bullets, no feature grids, no second caption under the headline.
One ground for the whole film (the template's world layer: a slow key light, vignette, grain) and one accent: the brand's, else the template's. Accent on the key word, the prompt and the result line only. Text >= 7:1 on the ground.

Music: a produced track, cut to the picture#

showtime audio cuts --for launch --apply <job>/project (or a track id from showtime audio music search): it picks the catalog track, fetches it once (announced, with its size), finds the excerpt that starts calm and swells where the end card starts, moves the scene changes onto its phrase starts (retime --cuts) and writes the excerpt into audio/mix.json. It prints the plan: excerpt, scene starts, how many land on phrases, the swell, the dynamics.
Run it again after any length or scene-count change; --offset keeps an excerpt you chose.
No sound effects by default. At most one soft effect (a low whoosh under the through), never one per cut. No voice unless asked; with a voice, duck the track under it (sound-design.md).
Credits are automatic: credits.txt, the block in share.txt and the end card's credit line. Keep the credit in the post's description (the composer's Content ID reads it).

Honesty#

Every word on screen traces to the README, CHANGELOG, site or docs, or to a run you saved: terminal text is copied from <job>/work/evidence/<name>.txt (command + exact output), never typed from memory. No speed claims, counts, users, stars or quotes that the sources do not state. The version and the install command are verbatim. A plausible detail (a backup file name, a flag, an output line) that nobody ran is an invented claim.

Steps#

Job. showtime job init <product>-launch --goal "<the request>" --platform <x|linkedin|youtube|reels>. Done when: SHOWTIME.md exists.
Source and evidence. Read the repo in the order of capture.md §2 (README, manifest, CHANGELOG, site). CLI or library: run each command you will show on a small sample input and save command + output to <job>/work/evidence/<name>.txt. Web product: showtime site capture <url or --serve dir> <job>/work/capture --aspect 16:9 (and 9:16 for a vertical cut). Brand: showtime brand init --from <repo> -o <job>/brand.json when the product has colours or fonts. Done when: you have the promise in the product's words, three results you can show from evidence or captures, the install/CTA line verbatim, and the colours and fonts.
Plan. Contract ("This video tells ___ that ___", story.md §1), then the scene table above filled in: per scene the words on screen, the evidence file or capture, and the handoff. Log it: showtime job note <job> --stage plan --verified "contract: ...". Done when: every scene has one message and a source, and there are 4-6 scenes.
Project. showtime new launch <job>/project --duration <len>. Replace every SLOT:: the hook (complete at frame 0; the key word in <em>, one of its letters with a closed counter in <span data-portal="counter">), the window (terminal lines from the evidence files; .hist repeats the previous command with its whole output or none, never only the last line, so it reads as one session), the headlines, the end card. Brand: override --bg --fg --muted --accent and the two font tokens at the top; fonts with showtime assets font "<family>" --copy-to <job>/project/fonts (components.md §7). Done when: grep SLOT finds nothing.
Music. showtime audio cuts --for launch --apply <job>/project (step "Music" above). Done when: the plan says the swell lands on the end card, or you chose another excerpt and said why.
First look, at every aspect you deliver. showtime check <job>/project and showtime snap <job>/project --every 1, then the same with --size 9:16 (and 1:1 when asked): the page re-lays itself per size, and a command that fits at 16:9 can be cut off at 9:16. Look at every sheet with the checklist below (text inside the frame and its box in every still, the hook still and large, no type sliding or shimmering). Show the user in one line and keep going (quick mode). Done when: 0 errors at every size (cut-off or off-frame text is an error), every WARN read.
Final. showtime render <job>/project --job <job> (16:9, final.mp4), and for a vertical cut also showtime render <job>/project --job <job> --size 9:16 (1080x1920.mp4: the same page re-laid; it checks the layout at that size first and stops when text would be cut off).
Verify. showtime qa <job> (and showtime qa <job>/1080x1920.mp4): quote the verdict, LUFS, true peak and the rhythm line. Publish-bound: showtime review-pack <job> and the critic, whose brief carries the checklist below.
Deliver. Write the post copy into share.txt above the credit block that render put there (1-3 sentences in the product's voice, only claims the video makes, the install line or link), exports on request (showtime deliver exports <final> --targets x,linkedin), the delivery card (modes.md §5).

Checklist (self-review and the critic's)#

[ ] 4-6 scenes; qa rhythm: layouts <= 6, 0-1 hard cuts, every scene change a move.
[ ] Frame 0 is the hook, complete and readable at phone size; it says what the product promises.
[ ] Each scene has one message and one proof, held >= 1.5 s after it lands; nothing is still for more than ~4 s.
[ ] One display face + one mono, sizes at or above the floors above, one accent, one ground.
[ ] Every camera move has a reason you can say (at most one fly-through, hook into product); otherwise the camera is still. No punch-ins, shake, bounce, pops or per-cut effects.
[ ] The music is a produced track with dynamics (qa music dynamics >= 3 dB); scene changes sit on its phrases; the name lands on the swell; it ends on a fade or a cadence, not a cut.
[ ] Every claim, command and output line traces to a source or an evidence file; the install/CTA line is verbatim.
[ ] End card: name, value, install/CTA, URL, credit line; held >= 3 s.
[ ] The 9:16 cut keeps every word inside the feed safe zone and nothing smaller than the floors.

Pitfalls#

A layout per feature. Three features are three commands in one window, not three designs.
The window left empty while a long command types: keep commands short and type them fast (data-fit 0.8-1.2 s); show the output within 3 s of the scene start.
A made-up UI presented as the product. If capture fails, stop and say so; a recreation is labelled.
Music chosen by name. Read showtime audio music info <id> (moods, energy, ending) and the cut plan's numbers; a track whose excerpt has no swell makes an end card with nothing to land on.
Retiming by hand after audio cuts: rerun it instead, or the cuts leave the phrases.