Ir al contenido

launch-kit: launch and update videos for your own product

Esta página aún no está traducida; aquí tienes la versión en inglés.

Use when a maker wants to announce a product or a new version: a short demo video of the real product, a looping GIF for the README, one short clip per feature for X / LinkedIn / Shorts / TikTok / 小红书, Product Hunt gallery stills and an OG image, post copy per platform in English and 中文, and a posting schedule. It is not for a talking-head promo of something you made (that is promo-recut) or a concept explainer (explainer).

Inputs → Outputs. A product (name, one-liner, site), release notes (a CHANGELOG section, a git range of Conventional Commits, or pasted notes) or a brief with the on-screen lines, the product itself to record (a URL or an Electron app, plus a shot list), or screenshots / recordings you already have (capture.clips), and the brand (logo, icon, colours, fonts) → kit/:

Folder What
demo/ demo-<lang>-16x9 / 9x16 / 1x1.mp4, 30–60 s: a product film with a varied rhythm (full-bleed UI with the camera pushing onto the action, 2.5–4× detail cuts, split screens, the window floating with a small lower third, short type beats between sequences, chapter cards, a logo end card), cut on the beats of a soft music bed; a .cover.jpg per video
readme/ loop-<lang>.mp4 + .gif, 15 s, seamless (the last frame crossfades into the first)
clips/ <feature>-<lang>-1x1 / 9x16.mp4, 10–20 s each, one feature + end card
stills/ ph-gallery-1-hero.png, ph-gallery-N-<feature>.png (1270×760), og-1200x630.png
copy/ COPY.md (Product Hunt fields, the launch post and every clip post, international platforms first), <platform>.<lang>.md
schedule/ SCHEDULE.md + schedule.json: launch day = the demo everywhere, then one feature clip per weekday
firstpass.json vstudio.firstpass on every video (+ <video>.firstpass.md)

Everything content-specific lives in ONE file, launch.config.yaml; start from $VSTUDIO/workflows/launch-kit/examples/launch.config.example.yaml, which documents every key.

Pipeline ($VSTUDIO = repo root; LK="python3 $VSTUDIO/workflows/launch-kit/scripts/launch_kit.py")

Section titled “Pipeline ($VSTUDIO = repo root; LK="python3 $VSTUDIO/workflows/launch-kit/scripts/launch_kit.py")”
  1. Draft the features from the release.
    Terminal window
    $LK features launch.config.yaml # release.changelog | release.git {repo, range} | release.notes
    vstudio.launch.features keeps feat / fix / perf / breaking commits (chores, ci, docs, tests, refactors never become features) or the bullets of a Keep-a-Changelog section (“Mac app:” style intro bullets contribute their nested bullets). Pick 3–6. For each, write ONE on-screen line (≤ ~50 characters in 16:9; the caption is fitted and wrapped by measuring the real font, and the build fails on a line that cannot fit) and mark the word to highlight with 【】. Captions say what the viewer gets, not what the code does.
  2. Record the product. Write shots.yaml (one shot per feature: the actions to perform, 5–12 s each) and run
    Terminal window
    $LK capture launch.config.yaml # = node $VSTUDIO/workflows/launch-kit/scripts/capture.mjs --shots shots.yaml --out capture ...
    capture.mjs drives the product with Playwright (a web URL in Chrome, or an Electron app with a throwaway profile), at 2× device scale, records the page with CDP screencast frames assembled at a constant 30 fps (not Playwright’s low-bitrate recorder), draws a smooth cursor with a click ripple, and types at a human pace. It writes capture/<shot>.mp4, a still per shot, and capture/shots.json with the focus boxes: where and when each click / typed field happened. An Electron app gets its throwaway profile through capture.env / capture.env_json (--env-json); examples/reelfold/ is a full example (Reelfold recording itself: seed.py builds a temp profile with real engine data on a synthetic talk, shots.yaml drives Home → plan → project → transcript → chat → publish). The camera uses them for the punch-ins. Seed the product with sample data that is safe to show (no real names, emails, handles or other people). Already have screenshots? List them under capture.images with a focus box; the camera then makes a slow move to it.
  3. Build + render.
    Terminal window
    $LK build launch.config.yaml && (cd kit/build/demo-en-16x9 && npx hyperframes check)
    $LK render launch.config.yaml [--only demo|loop|clip|<feature>] [--quality draft]
    One HyperFrames project per cut × language × canvas (kit/build/<name>/index.html, plan.json, cues.json). Layout per canvas keeps every caption inside the safe area of every platform that canvas is posted to (vstudio.platform safe boxes: 16:9 X / LinkedIn / YouTube / B站, 9:16 TikTok / Shorts / Reels / 小红书 / 抖音, 1:1). The look is a product film, not one template per feature (story.rhythm, film in the config). Features are grouped into short sequences; a sequence opens on a type beat (its first line, set large in Inter 560 / Noto Sans SC Medium on a soft gradient field, the words fading up out of a blur), its lead feature gets a full-bleed establishing shot (the UI fills the frame, no text, the camera gliding onto the action) and a detail cut (2.5–4× on the one element that changes, the take continuing across the cut), and the rest rotate through a floating window with a small lower-third caption, a split (two moments side by side / stacked, small labels) and full-bleed shots. A feature with chapter: opens a section on a chapter card; a first line that is the product’s name becomes a cold open + the name card. No two adjacent scenes share a layout, and words are on screen at most film.text (0.4) of the runtime: lower thirds, then type beats are dropped from the end until it fits. film.variety calm | balanced | lively, film.captions auto | type | lower | none, type.scale / type.weight, and per feature layout: full | macro | split | float pin it. Gradient fields only behind type; footage sits on a flat light field (brand.backdrop: studio = light, default; deep; paper - never black; a dark recording is lifted a little). Hand-overs follow the layouts: a hard cut into a detail, cut / push / rise between footage, a zoom-through out of a type beat, a blur dissolve into a card - each centred on a beat. film.arc: launch tells the product’s positioning as a launch film (studied from OpenFilm’s promo; see lib/vstudio/launch/story.py): a cold open on the product, a brand card (product.tagline as the kicker, the wordmark slammed in letter by letter with a spring, product.badges snapping in), a reel (one window, a hard cut on every beat through a moment of each feature, a filmstrip running under a fixed playhead, product.promise above it), then the chapters (chapter: per act; the title slams in; chips: [...] | platforms fan out under it), floating windows that swing in from a 3D angle and split panes angled like an open book, the camera eased in log space (a push reads even, the target never swims), a zoom-through from footage into each card, quiet code-made foley (swish per moving transition, tick per badge / chip / reel cut; motion.sfx: false = off), and an end card with the badges. film.blocks picks the opening (default brand card + reel): a waterfall (hundreds of real video thumbnails, film.waterfall.thumbs, pour down and land as the tiles of the wordmark, which resolves into the solid name in heavy type), a lift (the platforms’ own marks rise and lift “Create · Edit · Publish” / “做 · 剪 · 发” in heavy type) or a hook (one heavy line, for a short); brand.backdrop: cream is a warm light stage, type.heavy (700-900) the blocks’ weight and type.titles: heavy sets chapter titles in it too. A set from one config: films: [{id, shots, renders, <sections to override>}] + make: [films] renders every film to films/<id>-<lang>-<canvas>.mp4 with a cover and a contact sheet, and an index.html to review the set. Chinese is broken after its punctuation and subset to the characters on screen (Noto Sans SC Medium from ./install.sh; without it Regular with a hairline stroke), and an end card that animates the mark (brand.mark), the name, the call to action and the site. Delivery: BT.709 tags + faststart; loudness to −14 LUFS only when there is audio. Render one draft first and look at frames mid-scene and mid-transition before the full set.
  4. Stills, copy, schedule, check.
    Terminal window
    $LK stills launch.config.yaml # PH gallery + OG from the shot stills, rendered at 2x then resampled
    $LK copy launch.config.yaml # vstudio.publish.platform_post per platform x language -> copy/COPY.md
    $LK schedule launch.config.yaml [--apply]
    $LK check launch.config.yaml # firstpass, format "launch"
    Copy comes from the config’s post block (and features[].post), laid out per platform by the shared post engine (title fields, hashtag caps, X weighted 280, 小红书 / B站 conventions) and checked with vstudio.platform.check_text; without a post block the routed model writes it (llm task copy), and with no model routed the step fails instead of inventing text. --apply adds the schedule to the publish calendar (vstudio.project.pubcal) as planned posts for the accounts in schedule.accounts. Nothing is ever posted. $LK all launch.config.yaml runs steps 3–4.

From recordings you already have + a brief (no changelog)

Section titled “From recordings you already have + a brief (no changelog)”

Hand over the takes and a BRIEF.md and nothing is recorded: the config’s capture.clips lists the takes (id = the file name without its extension), and the config checkpoint is drafted by the engine before any step reads it (vstudio.project.adapters.launchkit): the brief’s shot table (a markdown table with the take file, a 0:04–0:08 range, the on-screen line in English and a 中文 column) becomes the features in order, one scene each (two files in a row = two shots under one caption, shown side by side in a split scene); the request sets the canvases (“16:9, 9:16 and 1:1” → demo.renders per language), the pace (“1.1×” → pace), the length, the transitions (“fast smooth animated transitions” → motion: {transition: mix, seconds: 0.3}: push (with motion blur), zoom-through and rise and hard cuts chosen from the layouts on both sides, ~9 frames, each centred on a beat of the music bed), the opening (“no hook” / “open on Home” → demo.title_card: false), the built-in music mood (“tech”) and the look (“dark” → the night theme, “teal” → the accent), and make: [demo] for a promo. A release source is only written when one is really there; without one, features lists the config’s features (or the brief’s table).

On a supplied take the camera finds the action itself (timing.motion: where the screen changes between low-res frames): an establishing frame, a fast eased push to 1.6-3× onto it so the UI text reads, a calm follow when the action moves, a slow creep to the cut (motion.camera: push, story.push_camera); a take longer than its scene plays the stretch that fills the scene at the cut’s pace, preferring the one that ends on the take’s result (the opening take plays from its first frame). Caches (idle spans, motion, posters) go under <out>/capture/, never next to her files.

A step that fails prints one plain line (launch-kit config: … → failure code stage-launch-config, or launch-kit <step> failed: …), never a Python traceback (VSTUDIO_DEBUG=1 shows it).

  • No voice by default; a license-safe music bed. music: auto (default) takes a built-in bed from the shared library (vstudio.music: generated in code, free for any use; mood tech for a product video). A mood name picks another built-in bed, none turns it off, and your own track needs music.file + music.license. A TTS voice-over (voiceover.enabled, one vo line per feature) is labelled “AI voice” on screen; say so in the post too. With audio the delivery is loudness-normalised to −14 LUFS.
  • Real product, real data you may show. Numbers in the gallery facts must be ones you can stand behind; sample data gets gallery.note (“Sample data”).
  • International platforms first, Chinese after, everywhere platforms are listed (copy, schedule, captions).
  • Red is not a brand accent here. brand.warnings flags a reddish accent and low ink/paper contrast.
  • Demo length: scenes are 4.5–9 s; the total is squeezed into demo.max_seconds (never under 4 s per scene) or padded by holding each shot’s last frame up to demo.min_seconds.

Real recordings wait: a plan being made, a render finishing. ffmpeg freezedetect (a very low noise threshold, so typing still counts as motion) finds the frozen stretches of each shot; they play at 4× after a 0.35 s lead-in at normal speed, the actions play at the feature’s rate. When the scene is still shorter than the ramped take, every segment speeds up evenly. Each segment is its own back-to-back clip in the composition. Detection is cached next to the shot (<shot>.mp4.idle.json); ramp: false on a feature turns it off.

The shot is cover-fitted into the product window. Each focus event becomes a punch-in (zoom ≤ the point where the acted-on element fills ~60 % of the window, ≤ 2.6×) that starts 0.55 s before the action, holds ≥ 1.5 s, and pulls back when the next action is far off. On a narrow window (9:16) the camera follows the action instead of showing the whole screen. Events closer than 0.6 s share one move. Translations are clamped so the shot always covers the window.

  • Record at 2× and extract video frames as PNG in the render (--video-frame-format png): UI text turns to mush with JPEG frame extraction or Playwright’s default recorder.
  • HyperFrames: never put a CSS transform on an element GSAP moves (gsap_css_transform_conflict); the camera’s first pose is a tl.set at 0. Overlay start states are opacity: 0 in CSS.
  • Media clips own their absolute data-start/data-duration. A scene containing timed videos is an untimed visual container animated by the main GSAP timeline; never put a second media clock on its parent.
  • Platform glyphs need staged local fonts too, including Chinese marks in English films. Do not rely on an undeclared system-font fallback. Clear lint errors before treating a browser layout check as executed.
  • A reel preview shows a result window at the requested pace, not a whole operation accelerated to fit a tiny slot. When a feature has multiple supplied views, the planner prefers exposure continuity without reordering features; the intended result remains preferred when exposures are similar. Posters cannot establish motion safety: rerender and check the final file, particularly dark examples surrounded by light UI.
  • Inspect what each recording actually shows. A filename such as “watch all” may contain a debug log rather than a creator-facing workflow. Use another verified source and narrow the claim, or omit the unsupported promise. Pair an actual editor/input view with its result when that explains the feature better than a lone dark example. Never manufacture a result or change its colours merely to make a check pass.
  • A loop needs its last frame to be its first: the last scene crossfades into a still of the first scene’s opening frame, and the first scene has no entrance animation.
  • Fit captions by measuring the actual font (PIL on the woff2), not by character counts: Newsreader and Noto Serif SC differ by 40 % in width at the same size.

The prepare stage links supplied media under the item’s inputs/ directory. In capture.clips and capture.images[].file, use those paths (for example inputs/01-home.mp4), not bare filenames relative to launch.config.yaml. The loader accepts an older bare filename only when that exact file exists in inputs/; explicit paths keep their meaning and genuinely missing files still stop before capture/render.