launch-kit

Launch kit — ship the media with the release

You are producing a RELEASE's media, not a video: one take of the shipped feature becomes every asset the release needs, composed, sized and verified per destination, kept as editable documents so the NEXT release is a re-render, not a re-shoot. Everything is data; the preview is the render; export is free at every resolution up to 4K, no watermark; the engine is MIT.

The kit is RENDERED in one verb, vos deliver, from documents that carry the taste: every card is a frame of a POSTER document (a plain take whose card sits where the poster wants it, with the words and the mark as clips), every cut plays the motion vos plan proposed into its document (an entrance, an end card, captions, a bed), the still moments come from the story, and two verifiers say what is wrong in words. deliver composes nothing. Your judgment goes into three files (BRAND.md, LAUNCH.md, actions.json), one pick (which poster on the shelf the cards follow) and one decision (which moments are the story), never into hand-cropping.

Setup

npm i -D @vosjs/cli    # the vos CLI (take pipeline included)

ffmpeg on PATH covers what the CLI does not emit (the PH hover-GIF) and lets vos validate --picture read a video's first and last frame. MP4 renders need system Chrome (--format mp4).

1. Establish the release

Five facts before any recording:

• What shipped — the feature, the version string, the URL where it runs.

• Which destinations — the per-channel spec table ships as data: node_modules/@vosjs/cli/schema/channel-specs.json (CWS, Product Hunt, X, LinkedIn, GitHub, OG, YouTube; sizes, counts, byte and duration ceilings, the word policy and the safe rect per destination, and each image's GENRE: screenshot is the real page, card is a frame of a poster document; references/channel-specs.md is the same data as a table). Loop over it, never hand-type dimensions. Ask which channels this release ships to; default to last release's set. The JSON carries a verified date; if it is more than a quarter old, spot-check the channel docs before shipping.

• The brand — resolve it BEFORE authoring any asset, and never default to a layout's own palette. The project folder's BRAND.md is the brand kit (frontmatter carries the colour roles bgA bgB bgC ink accent, the face roles fontDisplay fontBody, logoUrl (a bare mark, never a tiled icon), wordmark, and look, the presentation the site's own ground asks for: plate for a paper site, dark for a dark one, else gradient). Absent one, vos brand <url> witnesses it from the site's /design.md, /llms.txt and the page itself, and writes every role with its provenance; read it, correct what a page cannot say, and file it with vos recipe push BRAND.md --folder <slug>. Place it BESIDE the take (or in the take's parent folder): plan and deliver read it there with no flag.

• The words — LAUNCH.md beside the take carries the release's roles in its frontmatter, read with no flag:

  headline: "Two to six words\nover up to three lines"   # the posters and the end card
  kicker: "PRODUCT  V2.1"       # absent = the wordmark plus --release
  music: upbeat                 # a catalog slug or mood; none = silent
  entrance: tilt-in             # tilt-in | pull-out | rise | none
  endCard: on                   # the official End card template; none switches it off; a title, vos id or doc.json names another
  with: "Split cover, landscape@end"   # optional: any template laid at an anchor (@end | @start | @step:<id> | @<seconds>)
  captions: on                  # none switches the beat captions off
  poster: poster/landscape      # optional: the document every card renders from
  poster-tile: poster/tile      # optional: one class named on its own
  

  A headline is the ONE line the release says; write it over its lines with \n. vos plan reads the motion roles when it proposes the cut's motion and plan --style types the words into a poster; deliver reads only the poster roles, and with none it looks for poster/<class>/doc.json beside the take.

• The layout — a poster the cards will follow, on a shelf. The official Poster families project on vos.so holds one exemplar per aspect class (landscape, square, portrait, tile) beside its POSTER.md, which says when each class is the one; a maker's own project is the same shape. vos folder pull poster-families lists them with their vos ids. There is no layout name and no --layout flag: the exemplar IS the layout, copied by data.

The destinations decide the VIEWPORT, before anything records: footage resolution = viewport, and a 1280×720 take cannot honestly fill a 1920×1080 video spec. Record at 1920×1080 for a 1080p kit, 2560×1440 when a destination is larger.

If the work lands in a vos.so project (folder), pull it first and read every .md recipe in it — LAUNCH.md binds this loop the way CUT.md binds a cut, and a POSTER.md binds the posters. Recipes override this skill's defaults.

2. Source: one take of the real thing

The kit is made FROM the product, never from a mockup (store policy agrees: misleading listing images are a removal-grade violation).

Stage the content like a set before recording. Half of what separates a premium launch image from a screen grab is what is ON the screen. The actions.json must leave the product in the state a proud screenshot would show — labels typed, real-looking data, the feature mid-story — before any poster or store still is cut. deliver drops a blank moment (a wallpaper, an empty canvas) and says so, but it cannot stage the set for you.

Write the story into actions.json. Give steps an id (the moments and the re-render loop address them by it) and a caption where a beat deserves one (two to eight words; plan proposes it as a lower-third at that step's moment on the cuts that take words). After a click that navigates, put a wait for the load: the still is read at the END of that wait, so the frame shows the page, never the spinner.

• Fresh recording: the product-video skill's loop (explore → actions.json → vos record --strict → tune doc.json). Keep actions.json in the repo — it is the next release's script.
• Existing take: cut it with the vos-cut skill. A hosted take comes home with vos fetch <vosId|watch-url> --out dir --media.
• New version of a shot product: re-record with vos record into the SAME take (the footage is replaced, the cut survives as doc.prev.json), then vos plan take --reuse re-times the cut onto the new recording and names what could not follow. Never start from scratch; every fix is an edit to doc.json.

A motion-graphic segment rendered elsewhere is an INPUT: it drops in as a media overlay clip or a backdrop in the document, never the other way round.

3. The posters and the motion, as documents

The cut moves by data. A fresh vos plan proposes the cut's motion into doc.json from LAUNCH.md's roles: the card's entrance, the END CARD (the official End card template on vos.so, laid at the end: the card recedes over a one-second freeze of its last frame, then the mark, the headline arriving word by word, the release line and the URL settle on the ground; its clips are stamped from: endcard), any template with: names at its anchor, a caption per step, a music bed and a click sound on every press when the take has no mic; every proposal carries a stable id (bed, click-<n>, caption-<step>). On an existing cut --motion re-proposes; a refresh never does, so a deleted end card stays deleted. A template is a plain take on a shelf whose clips carry stable ids; a ref is a title on the official shelf ("End card", "Split cover, square"), a vos id, or a document on disk. Open the take in the studio and what you see is what the kit renders.

vos plan take --motion --release v2.1

A poster is a DOCUMENT, a plain take: the card placed by frame.inset (a negative side bleeds it off the edge), leaned by the tilt track's rest pose, the words and the mark as overlay clips with stable ids (stage-title, stage-kicker, stage-brand, stage-mark), and a trailing hold on the last segment whose START is the still. Make one per aspect class the release needs (landscape, square, portrait, tile), each from the take's own footage:

cp -r take poster/landscape            # or vos duplicate <take vosId>, then fetch --media
vos plan poster/landscape --style <poster vosId|doc.json> --headline "…" --kicker "…"
vos frames poster/landscape --frame <rest>; vos validate poster/landscape
vos push poster/landscape --folder <project-slug> --label "poster, landscape"

plan --style <poster> copies the LAYOUT (the card's placement and presentation, the stage clips with the release's words patched in from LAUNCH.md or the flags, the rest lean, the hold) onto the take and keeps the take's own aspect, chrome and cut; the brand's mark from BRAND.md logoUrl is fetched into brand/ for stage-mark, and what could not follow is said. Narrow the segments to the moment (a zoom apex, the settled response after a click) so the rest is the feature, never the cold open; the product may keep PLAYING inside the card until the hold. Then look at the rest frame: a poster is judged by eye before it is pushed. A human opens it in the studio and drags the card, retypes the words in place, moves the hold; vos pull --media brings that back down, and the next release re-words it with the same plan --style from its own pushed document.

4. Deliver

vos deliver take --to cws,producthunt,x,linkedin,og,github,youtube --release v2.1

One pass, and the verb decides what you used to decide by hand:

• The moments. Still times come from the STORY: every step's end plus a settle (the end of the wait that follows a click), then the zoom apexes, then a spread. Each candidate is read once as the real page; blanks are dropped, two of one frame collapse to one, every drop said in skipped[]. --times step:<id>[+offset] names one.
• Screenshots are the real page, full bleed: the store still and the gallery drop the cut's zoom, tilt and chrome by default (--composed keeps them, and the picture pass refuses it for the store).
• The cards render from the posters. Each card destination (the OG card, the LinkedIn and X images, the YouTube thumbnail, the CWS tile and marquee, the GitHub social preview, the PH thumbnail) renders from the poster document of its aspect class beside the take (poster/<class>/doc.json, or the document LAUNCH.md's poster roles name) at that document's rest, at the destination's exact pixels. A class with no document is the take's own frame, said once in skipped[]. Nothing is composed in memory.
• The cuts play the document. deliver keeps only each destination's MECHANICS: a loop drops the entrance, the end card, the captions and the sound; a silent channel mutes the bed; the 9:16 cut is a reframe whose crop follows the camera, not a letterbox; a loop destination the take outruns takes the take's first seconds up to its cap; a byte ceiling becomes a bitrate budget.
• The manifest. kit.json beside the assets records every asset with its destination, the moment it came from, and for every card the poster it is a frame of (source: "poster", the class, the file, the vos it tracks, the shot rect and the text boxes read FROM the document), and skipped[] with every reason.

5. Verify: the specs, then the picture

vos validate kit/kit.json --picture

The spec pass re-measures every asset from its bytes (px, bytes, duration, count, a WebP under a .png name). The picture pass says what each asset LOOKS like, every finding with a code, a severity, a fix hint and a box:

code	what fires
blank	a card whose subject is under the ink floor: a wallpaper, an empty canvas
duplicate	cards of one frame (two is a note; three or more fails)
subject	a card off the 60 to 92% band, or a crop where a card was asked for
separation	a light card on a light ground with no shadow and no drawn edge
halfsize	a tile that loses its edges when the store shrinks it
sliced	a headline crossing the frame edge
safe	words outside the destination's safe rect, or words where none are wanted
contrast	a text box under APCA Lc 60 (headline) or 75 (body)
firstlast	a cut that opens or ends on nothing, or bled on all four sides

A problem is redone by fixing the INPUT (a moment, a word, the set, the poster document, a recipe line), never by hand-editing a PNG. Self-check by these names before you render: a cold-open hero is blank, eight crops of one frame is duplicate, white words on a paper ground is contrast. A poster's card is judged by the shot rect the document carries, so a card you dragged off the band is subject before anyone posts it.

6. Judge beside the references

vos judge kit/kit.json --against <MANIFEST.json>

The manifest names the maker's reference set (a private folder: id, file, role, layout, facts, rule per asset; the public evals reference it by role). For every still with a reference of its role the verb writes two sheets (the asset left, then right) and the rubric beside them, and leaves judge.json with a slot per pair. Judge every pair BOTH ways, with the rubric's numbered rules and the three positive tests (would you post it as a still; does it read at half size; name three ways it acknowledges THIS product), and write win true, false or null (a tie) with the rule numbers. Parity with the references is 50%; a kit under 40% is not ready to market, said in the handoff, never shipped around.

Skip the judge for a re-render whose inputs did not change (the same take, the same words); run it when a layout, a look, a headline or the set of moments changed. Do not re-run validate or judge after a push unless the human asks.

7. Push the release, file the kit

Push the source and every poster labelled for the release, and FILE the kit's stills into the release's project so the human can retrospect without a terminal:

vos push take --folder <project-slug> --label "v2.1 launch" --note "<what shipped, one line>"
vos push poster/landscape --folder <project-slug> --label "v2.1 poster, landscape"
vos asset push kit/*.png --folder <project-slug>

End by handing the human the loop, not the files: the watch page plays the latest version, the studio edits the cut AND the posters (the card is a layer you drag; the words are retyped in place), and vos pull --media brings their edits back down. Next release, start at step 2's third bullet and re-word the posters with plan --style from their own pushed documents.

Preserve the human's changes. They edit the take in the studio outside this conversation. If vos pull --check or the differ's protected set shows a change you did not make, assume it was intentional or ask; never overwrite it, and never re-plan over a manual span.

Launch week (5-12 clips, one style)

A launch week is a series: cut and sign off ONE seed clip with the human first, then cut every other feature's take with vos plan <take> --style <seed doc.json|vosId> so the batch shares its look by data. Never spread before the seed is signed off.

Notes

• deliver is the procedure; this skill is the judgment around it (the set, the words, the layout pick, the moments, the verdicts). If you find yourself cropping a PNG by hand, the fix is a document, a recipe line or a check, and you should say so in the handoff.
• Platform specs drift. The JSON carries a verified date; if it is more than a quarter old, spot-check the channel docs before shipping.
• A layout is a document on a shelf plus a recipe line, never a name the document learns: the official Templates project holds the split cover in four aspect classes and the end card today (vos plan take --style "Split cover, landscape" carries a layout; a poster's still is where its trailing freeze begins); a maker's own family is a project of their own posters beside a POSTER.md. Applying one is plan --style <vosId>.

Avoid (the traps that shipped)

• A 720p recording against a 1080p spec: the destinations pick the viewport before anything records. Cost a full re-record once.
• A layout's own palette on a deliverable: resolve BRAND.md (or witness the site) before any asset is authored, and put it beside the take.
• A cold-open hero (blank): the first frame is the marketing page's headline with the product nowhere; deliver reads the steps, but a script with no gestures gives it nothing to read.
• A poster whose rest is the cold open: narrow its segments to the moment (a zoom apex) and hold there; the still is the hold's start.
• A card that is not a frame of a document: there is no --poster, no --shot-time and no in-memory composition; a card destination with no poster document renders the take's own frame and says so.
• Eight crops of one frame (duplicate): a script with one moment. Give the story steps, and waits after the clicks that navigate.
• A store screenshot under the frame chrome or the camera zoom: the store still is the real page, full bleed (deliver's default, and --composed is refused by the picture pass).
• The loading plane as the hero: a click that navigates settles when the wait after it ends; a click with no wait after it is read 0.4 s later, mid-load.
• White words on a paper ground (contrast): the end card takes the brand's ink; a kicker softened too far reads under the floor.
• A card that is WebP under a .png name: vos still writes WebP; convert, then vos validate reads the bytes and says so.
• Padding a spec floor: a 36 s story is a skipped 60 s demo, with its reason.
• A hand-typed dimension: every size comes from channel-specs.json.
• The save beat on a demo instance that disables writes: say so in skipped, never fake the click.