In film, the second unit is the crew that shoots everything the principal unit doesn't: plates, inserts, establishing shots, product close-ups, b-roll. No stars, no dialogue — just the footage the film can't be cut without.
A working studio for making product demos, marketing videos, and AI-generated character film — the recording engine, the cutting pipeline, and a set of Claude skills holding the prompt craft for image and video generation.
It came out of shipping 30+ videos for one product: feature demos filmed from real UI, social cuts derived from them, and a mascot character with a locked look and a cast voice. Everything product-specific has been stripped; what's left is the method and the tooling.
Two halves, usable independently:
| Real UI — record an app you actually run, verify it, cut it for each platform | scripts/, templates/, examples/ |
| Generated — characters, locations, props, scenes, video prompts, voice, lip sync | skills/ |
The line between them is deliberate: the interface is always real footage; generation is for everything around it. Fabricated UI reads as fake to viewers who can't say why.
git clone <your-fork> second-unit && cd second-unit
cp config.example.env config.env
scripts/bootstrap.sh # toolchain check, deps, Chromium, library dirs
node scripts/demo-app.js & # a small bundled app on :5050
node examples/01-record-desktop.js # records it — no configuration needed
node scripts/qa-take.js out/desktop/*.webmThat works on a fresh clone with nothing configured. demo-app/ is a deliberately small
fake app that exists so you find out whether ffmpeg, Chromium and the cutter work on your
machine before you also debug your own app's selectors. Keep it around: when a recording
breaks later, running the same recorder against the demo tells you in a minute whether the
problem is your app, your selectors, or the studio.
Then point it at your own product — config.env is the only file you should need to touch:
APP_URL="http://localhost:3000"
APP_READY_SELECTOR="[data-testid=app-root]"Nothing in scripts/, lib/ or templates/ hardcodes a URL, a selector, or a brand colour;
node scripts/check.js enforces that.
Requirements: Node 20+, ffmpeg, and a browser Playwright can drive. Python 3 with
pillow + numpy (pip install -r requirements.txt) only for the image compositors.
Four recorders, in increasing order of cost and fidelity. Start at the top; most demos never need to go further down.
| Recorder | Output | Cost | Use for |
|---|---|---|---|
examples/01-record-desktop.js |
1280×800 | fast | most feature demos |
examples/02-record-mobile.js |
540×960 | fast | source for vertical/social |
examples/03-record-4k.js |
3840×2400 / 2160×3840 | slow | hero footage |
examples/04-record-stopmotion.js |
same, sharper | ~15 s wall per 1 s output | the sharpest thing here |
examples/05-record-device-stage.js |
app live on a photoreal device | slow | hardware shots |
Two non-obvious things they solve:
You cannot get a 4K web recording by asking for one. Playwright's recordVideo captures
CSS pixels, and deviceScaleFactor never reaches the video stream. scripts/rec4k-common.js
makes the CSS viewport itself 4K and zooms the page — rewriting @media breakpoint values by
the same factor so responsive logic still behaves — and the layout genuinely re-rasterises at
3–4×. scripts/stopmo-common.js goes further: page.screenshot() does respect
deviceScaleFactor, so it drives the UI in deterministic steps and screenshots every output
frame. Native layout, native breakpoints, genuinely crisp text.
A UI pasted onto a photo of a phone never survives a second look.
scripts/device-stage.js renders the device frame and the live app together in one page, so
geometry is exact by construction and the screen's own glow spills onto the frame for real.
Measure the screen rect once with scripts/measure-frame.py; check geometry with
scripts/probe-stage.js before spending a take on it.
scripts/build-flowcut.js turns a raw take into a continuous condensed cut: no mid-action
cuts, text overlaid on the footage, per-segment speed (fast through typing and waits, 1× at
the payoff), landscape or vertical.
The thing that makes it work: recorders print MARK <name> <t>s at every beat, and segments
are cut on those marks. Cutting blind against a stopwatch lands cuts mid-keystroke.
node scripts/build-flowcut.js raw.webm out.mp4 spec.json [vertical]
node scripts/add-music.sh out.mp4 track.mp3 final.mp4 -14node scripts/qa-take.js <clip.mp4> # screen recording (default)
node scripts/qa-take.js <clip.mp4> --plate # POV / device-stage takeNon-zero exit on FAIL. Checks for dead captures, frozen holds, edge gaps where a plate under-covers the frame, dead tails, and held empty screens. Every one exists because that defect actually shipped once. Run it before you look at the video, not after.
The two modes matter, because "static" means opposite things. In a screen recording the app holds still while the viewer reads — that's the content. In a POV take the plate carries continuous hand-held drift, so every frame should differ, and a static run is the signature bug: a hold emitted as one long frame, which freezes the picture and makes the plate jump. Running the wrong mode gives you confident nonsense in both directions.
Markdown skills for Claude (or any agent that reads SKILL.md files). They're the accumulated
prompt craft — including the failures, which is most of the value. Drop the ones you want into
~/.claude/skills/ or your project's .claude/skills/.
Characters and assets
story-bible-builder— interview-driven: turns a story or brand into one dense canon document, shipped as an installable skill so every later prompt already knows the world. Start here if you're building a character.ghost-mannequin-character-sheet— the 3-panel reference sheet (headless front / back / face close-up) that locks identity separately from wardrobe.banana-pro-director-2.0— the image prompt grammar. Face locks, outfits, character sheets, scene plates, detail shots, outfit replacement.location-plate-builder— environments built as functional sets, not pretty backdrops.prop-sheet-builder— objects and floating UI/graphic elements.scene-prompt-builder— putting a locked character in a locked location without slop.
Motion and sound
cinema-worldbuilder-pro-30— the video prompt grammar. Includes the two mistakes that cost whole batches: silently rendering at 720p, and untimed multi-beat shots.character-voiceover— casting method, and why "robotic" is almost always rhythm rather than timbre.character-lipsync— generate with the audio rather than repainting the mouth after; what works, what silently no-ops.
Product
product-demo-videos— MARK-based cutting, per-platform specs, the hook rule.device-screen-studio— real UI on a real-looking device, without compositing artifacts.pov-device-film— the full runbook for a character using a device with the live app on it.
config.example.env the only file you should need to edit
lib/ config loader + the app-target abstraction
scripts/ recording engines, cutting, QA, compositors
templates/ device stage, card + caption renderers
examples/ five worked recorders — copy one per flow
demo-app/ a small fake app, so a fresh clone can record something
skills/ the SKILL.md pack
docs/ the recording guide
.claude/skills/ the /feature-video pipeline skill
node scripts/check.js is the repo self-check — syntax, local requires, doc path
references, brand neutrality, secrets, config hygiene. It needs no ffmpeg and no browser, so
it runs anywhere. It's what CI runs.
Everything app-specific lives in config.env:
APP_URL="http://localhost:3000"
APP_READY_SELECTOR="[data-testid=app-root]" # or APP_READY_FUNCTION for a global
APP_HIDE_CSS=".cookie-banner{display:none!important}"
BRAND_ACCENT="#4F46E5"If your app needs more than a URL and a readiness gate — an entry function to call, a fixture
user to seed — create app.config.js at the repo root exporting { async prepare(page) {} }.
It runs after the readiness gate, before filming.
- Only real UI. Generation is for atmosphere, characters, and hardware — never the interface, never text the product would render.
- Probe, then record. Never script clicks blind.
- Verify frames.
qa-take.jsfirst, then look at extracted frames. A statistic and a visibly broken composite are regularly true at the same time. - Flush immediately. Every verified video goes straight to the library.
- Captions or it isn't a post. Short-form is watched muted; an uncaptioned cut is a silent film of a UI.
MIT — see LICENSE. The bundled font is separately licensed; see assets/fonts/README.md.