product-demo-video
Skill0 starsChanged 36 days agoMITView on GitHub
Creates product demo, launch, and promo videos rendered from code, tuned for Claude Opus 5.5: checks the model, asks a minimal set of intake questions (including an optional AI voiceover with ElevenLabs, OpenAI, Google or Azure), captures the real product (its code with permission, the live web app as sharp vector UI, iOS and Android screens), plans shots on a beat grid, builds a deterministic seek(t) canvas animation in the viral Opus 5.5 style (one continuous take, mask-rise type, 3D UI with screen-studio camera, flood and iris transitions, count-ups, mascot), analyzes reference videos or X posts, critiques its own frames and motion until every score is 8+, renders MP4s in 16:9, 9:16 and 1:1 with music, SFX, voice and captions, and ends with a token and cost report. Use when the user asks for a product demo video, launch video, app promo, walkthrough video, product reel, motion ad, or animated explainer.
- Reads credentials
- Installs packages
5 other projects use a version of this.
Adopted by
Most recently taken up by
- EvoLinkAI/evolink-agent-skillstheir own version, first seen Nov 8, 2026
- HamedMP/matrix-ostheir own version, first seen Nov 2, 2026
- mikeOnBreeze/cc-crossbeamtheir own version, first seen Oct 8, 2026
- iart-ai/ecommerce-video-skillstheir own version, first seen Jun 22, 2026
- realruian/product-demo-videotheir own version, first seen Jun 9, 2026
Other versions5 projects use a version of this.
Word for word · 0
None found.
Edited · 0
None found.
Wrote their own version · 5
- mikeOnBreeze/cc-crossbeam★ 297same name
- HamedMP/matrix-os★ 86same name
- iart-ai/ecommerce-video-skills★ 17same name
- EvoLinkAI/evolink-agent-skills★ 4same name
- realruian/product-demo-video★ 1same name
How matches are found. Word for word: the exact same text. Edited: shares passages with this file (“99% same” means 99% of its five-word phrases also appear here; 10% or more counts). Their own version: similar meaning, or, weakest, the same name. We can’t tell which project had it first. Projects are counted among the 200 most-starred candidates, and mentions are links to the repository we found, up to 50 per site, so both counts are a floor.
In this repositoryOne other file it gives its agents.
- AGENTS.mdAGENTS.md
The file
Show its rules and text
We split this skill into the rules it gives an agent
18 rules. 0 are common, 8 are distinctive, and 10 are very specific to this project.
0
Common
Most of the 3 edited versions we read keep these.
8
Distinctive
This author's own choices. Worth a look.
10
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 47
Never block: if they continue, use the same workflow with no shortcuts.
- Distinctive: line 89
No voiceover (music and sound effects only)
- Very specific: line 91
If they pick a provider without a key: tell them to add KEY_NAME=...
- Distinctive: line 97
If the user says "just go", use every default (no voiceover) and infer features, metric and CTA from the site.
- Very specific: line 102
The user's folder gets the final MP4s and nothing else: Never create a project folder, node_modules, screenshots, drafts or config files …
- Very specific: line 122
Read house-rules.md in the workspace now and follow it for everything below.
- Very specific: line 136
Use its CSS variables / Tailwind / iOS / Android colors for THEMES (exact hex and radius), its fonts (@font-face from assets/code/fonts, or …
- Distinctive: line 186
Never click anything that writes data (save, delete, send, publish, logout).
- Distinctive: line 192
Read research.md: the measured grammar of the top Opus 5.5 videos.
- Very specific: line 202
No reference but want one?
- Distinctive: line 205
Write the story device first (top of briefs.md): logline, motif, world, hero UI moments, pace, carrier object, payoff.
- Very specific: line 225
Write docs/shotlist.md on the beat grid.
- Very specific: line 262
Keep window.seek = seekWith(draw) so frames load before capture.
- Very specific: line 289
Use sp(t, 'snappy'|'base'|'heavy'|'playful') for entrances and track() for anything with several targets (cursor).
- Very specific: line 339
After the full render, if you can start a separate agent or subagent, have a fresh one that didn't build the film run the critique pass on …
- Very specific: line 370
Only if they can't, use --transcript <chat.jsonl> and call the result a lower-bound estimate.
- Distinctive: line 391
No Math.random, timers, CSS transitions or carried state in render mode.
- Distinctive: line 397
No film without a story device: a motif that opens it and returns, and a payoff line.
“Common” is measured: kept by at least 6 in 10 of the 3 edited versions of this file we could read. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Product Demo Video (Opus 5.5 motion studio)
The model writes a program, not a video: index.html exposes window.seek(t), headless Chrome
paints every frame, ffmpeg encodes. The prompt is 10% of the result; the harness (real assets,
springs, beat grid, critique loop) is the other 90%.
SKILL_DIR below means the folder containing this file (e.g. ~/.claude/skills/product-demo-video,
~/.cursor/skills/product-demo-video, ~/.codex/skills/product-demo-video, or skills/product-demo-video
in a cloned repo).
Asking questions. Whenever this skill says "ask", use your structured question tool if you
have one (Cursor AskQuestion, Claude Code AskUserQuestion, ...). Otherwise print the options
as a numbered list, first option = recommended, and wait for the reply. Ask each round in one turn.
Step 0a: Model check (gentle reminder)
You know which model you are. If it is Claude Opus 5.5, say nothing and go to Step 0b.
Otherwise, before anything else, show this once (fill in your model name) and ask:
Quick heads-up: this skill is tuned for Claude Opus 5.5. You're on
<your model>. It will still work, but Opus 5.5 is noticeably better at the motion design, layout and self-critique this needs. To switch: Claude Code/model→ Opus 5.5 · Cursor: the model picker under the chat box · Codex / other agents: pick a Claude Opus 5.5 model in settings or your API provider (Anthropic, Bedrock, Vertex, OpenRouter). Use high thinking effort (max for flagship launches).
Question model: "Continue on this model or switch first?"
- Switch to Opus 5.5 first (Recommended): stop here; the user re-runs the skill after switching.
- Continue with
<your model>.
Never block: if they continue, use the same workflow with no shortcuts.
Step 0b: Toolchain and voice check (silent)
node SKILL_DIR/studio/doctor.mjsIt checks node, ffmpeg, the browser (via Playwright), the optional iOS
Simulator and adb, and which voice API keys exist (names only, never values) in
the environment, ~/.config/product-demo-video/.env, or a .env in the user's folder. Run it
here just to read the voice providers; Step 1 runs it again in the workspace. Fix missing tools
yourself where you can; tell the user only what needs them (e.g. installing ffmpeg).
Step 0c: Intake (the minimal questions)
Skip anything the user already answered. In plain text, ask only for what can't be multiple choice:
What's the product URL? Optional: the 3 features to show, one proof number, and the CTA. If the app is behind a login, I'll open a browser window and you sign in yourself; I only look, never change data, and only the session is kept until the video is done.
Then ask these in one round (first option = recommended / default):
| id | Prompt | Options |
style | What kind of demo? | Launch reel: hook, UI in action, 3 features, proof, CTA (Recommended) · Product walkthrough: the real app in a phone or Mac, one chapter per feature, callouts · Story film: a motif or character carries the product from problem to payoff · Explainer: how it works, one idea at a time, with voiceover · UI morph loop: one shape morphs through 8-12 real UI states · Flagship film, 45 s+ |
length | How long? | 20-30 s (Recommended) · 15 s · 45-60 s |
format | Which formats? | 16:9 for web and decks (Recommended for walkthroughs) · 9:16 first, then 1:1 and 16:9 (Recommended for social) · 9:16 only · 1:1 only |
music | Music? | Original track synthesized in code + UI sounds (Recommended) · I'll supply a track (any audio file; copy it to audio/track.wav in the workspace) · No music |
voice | Add an AI voiceover? | built from doctor output, see below |
look | Which look? | Keynote continuous take in the product's brand: one unbroken shot, UI in 3D, flood and iris transitions (Recommended) · Kinetic launch: word slams, chapter labels, hard cuts on the beat · Match a reference video or X post I'll give you · Both brand and neutral house style: two versions |
mascot | Add a mascot character? | No (Recommended) · Yes: a simple box character that reacts on the beat (or describe yours) |
code | Can I read your product's code to get its exact colors, fonts, icons, wording and screens? Read-only: I never run, build or change it, and never open .env or key files. | Yes, it's this folder (Recommended when this folder has the product's package.json, .git, pubspec.yaml or Xcode project) · Yes, it's in another folder (I'll give the path) · No, use the live product only |
platform | Where does the product run? (skip when the URL makes it obvious) | Web app or website (Recommended) · iOS app: I'll open it in the iOS Simulator · Android app: emulator, or a phone with USB debugging · Web and mobile |
voice options, always in this order. Mark each (key found) or (needs KEY_NAME in .env)
from the doctor output:
- ElevenLabs, most natural voices (Recommended)
- OpenAI TTS
- Google Cloud TTS
- Azure AI Speech
- Free local voice (macOS
say/ espeak), for drafts and timing only - No voiceover (music and sound effects only)
If they pick a provider without a key: tell them to add KEY_NAME=... to
~/.config/product-demo-video/.env (works for every future video) or to .env in their own
folder (template: SKILL_DIR/studio/env.example). Never ask them to paste a key in chat.
Offer to continue meanwhile with the free local voice for timing and swap the real voice in
before the final render.
If the user says "just go", use every default (no voiceover) and infer features, metric and CTA from the site.
Step 1: Workspace
The user's folder gets the final MP4s and nothing else. Never create a project folder,
node_modules, screenshots, drafts or config files there. All work happens in a workspace
outside it, named after the product:
node SKILL_DIR/studio/workspace.mjs init <product> # run from the user's folder; prints WORK and ROOT
cd <WORK> # every command below runs here
node doctor.mjs # must be all ✓ for required tools
# only for a supplied music track: pip install numpy librosa soundfileinit creates ~/.cache/product-demo-video/projects/<product>/ with the studio files, records the
user's folder as ROOT in .pdv.json, and installs Playwright (the driver) once for every future
video; it drives Playwright's headless shell if present, else the Chrome or Edge already installed, and
downloads Chromium only if none is. Run the studio scripts with full permissions (outside any agent
sandbox): a sandboxed browser can't start, and the scripts say so. If the
workspace already exists (a follow-up edit), it is reused: the film (index.html), assets, audio
and docs stay as they are, and the studio tools, kit and house rules are refreshed to the current skill. Remember ROOT: Step 8 copies the
final MP4s there.
Read house-rules.md in the workspace now and follow it for everything below. Note the start
time; you'll report total time at the end.
Step 2: Capture real assets
Take the best source available, in this order. The camera pushes in 2-3x: a screenshot goes soft there, while vector UI stays sharp and every card, row and icon can move on its own.
1. The product's code (only if the user said yes to code):
node scan-code.mjs [path] # default: the user's folder -> assets/code/code.json + svg/ fonts/ icons/ copy/Use its CSS variables / Tailwind / iOS / Android colors for THEMES (exact hex and radius), its
fonts (@font-face from assets/code/fonts, or the next/font names from Google Fonts), the logo
from assets/code/svg, screens to pick shots, and copy/ for real wording. Then read the files
under read for the screens in the shot list: they show the real layout, states and labels to
rebuild with card, chip and toggle. Read-only: never run, build, install or edit the repo,
never open .env or key files, and never put source code in the film. If the user already runs
the app locally, ask before using that dev server; then capture it like a site (step 4).
2. Site and brand: node capture-site.mjs https://product.com (hero, sections, brand.json).
3. Logged-in app: say "A browser window will open: sign in there and I'll take it from there,
only looking", then run node capture-app.mjs --login [/paths]. SSO, 2FA and passkeys work, no
password goes in a file, and only the session is kept (.auth.json, deleted by workspace.mjs clean).
APP_EMAIL / APP_PASSWORD in .env without --login remains for shared test accounts.
4. Vector UI of every screen you'll show:
node capture-dom.mjs /dashboard /leads # or full URLs; reuses the --login sessionOpen each assets/dom/<name>.compare.png (screenshot | vector | red = differences). At 90%+ use
the vector page; where it's red (canvas charts it couldn't see, masks, script-driven text), use the
screenshot for that part. It prints the CSS features canvas can't reproduce. In index.html:
const DOM_PAGES = { dash: 'assets/dom/dashboard.json' }, add loadDom(DOM_PAGES) to
window.ready, then domPage('dash', x, y, w, { region, each }) draws it (region = one card,
each(node) returns { a, dx, dy } to stagger or slide parts), domFind('dash', 'Revenue', x, y, w)
gives cursor and ring targets, and domCanvas('dash') is the texture for plane3d.
5. Native mobile apps (platform iOS / Android): node capture-mobile.mjs devices. Ask the
user to open the app in the iOS Simulator (or emulator / USB phone) and go to each screen, or jump
there with capture-mobile.mjs open <deep-link>; never tap or type yourself. shot <name> per
screen (clean 9:41 status bar), record <name> --seconds 8 for a real interaction (then
prep-footage.mjs). Show them in phone(x, y, w, drawScreen, { aspect }).
6. Your own browser tool (Cursor's browser, Claude in Chrome, a Playwright MCP), only when the
login window in step 3 can't sign in (a passkey bound to the user's own browser) and the user is
signed in there: navigate and screenshot read-only. For vectors, evaluate lib/dom-extract.js,
then JSON.stringify(__pdvExtract()), and save it as assets/dom/<name>.json; raster parts and
fonts are missing that way, so compare it with a screenshot first.
Screenshots still have uses: capture-parts.mjs /dashboard crops cards, and
probe-targets.mjs targets.json finds text boxes on screenshot-based shots. Write
assets/README.md listing every file and what it shows. Never invent screens or numbers.
Capture gotchas:
networkidlehangs on sites with live trackers; the scripts useloadplus a delay.- Apps often scroll inside an inner pane, so
fullPageadds nothing. Use a tall viewport (--h 3200) and crop parts from DOM boxes. - Off-canvas drawers (x > viewport) break clips and add shadows; the scripts skip them.
- Never click anything that writes data (save, delete, send, publish, logout).
- Entrance animations are jumped to their end state before a DOM capture; endless ones (marquees) stop where they are. Script-driven effects (text scrambles) can still differ: check compare.png.
Step 3: Reference, brief, style guide, script, beat grid
- Read research.md: the measured grammar of the top Opus 5.5 videos. It is the default look; every rule below assumes it.
- Reference (only for "Match a reference"):
node ref.mjs <x.com status URL | .mp4 URL | file>. It downloads the video (and the post text, which often holds the creator's full prompt), and writesrefs/<name>/sheet.png(2 frames per second), 12 fps motion strips of the opening and the busiest seconds (refs/<name>/moves/), cuts, shot lengths, palette, loudness and a motion audit torefs/<name>/analysis.md. Open the sheet and every strip, and look at them: the strips show how things move (what blurs, what carries across), which the sheet can't. It also takes a prompt-motion.com entry link and then saves the creator's full prompt torefs/<name>/prompt.md: read it for the techniques, timings and bans. No reference but want one?node gallery.mjs [words] --tag "Product UI"(orPhone,Kinetic type,Charts) searches prompt-motion.com; whatships.com is a second library. Pick one whose grammar fits the product. - Write the story device first (top of briefs.md): logline, motif, world, hero
UI moments, pace, carrier object, payoff. Then pick the template: E keynote continuous take
(the default
look), I device chapter walkthrough (every walkthrough), J story film (storystyle, or a mascot), K object chain launch (launch reels with one strong object or number), L keynote sting (10-15 s teaser, cuts at peak blur), F agent / chat product (the user types a request, the product does the work), G inside the period (looping; products that work while the user is away), H explainer (how it works), A kinetic launch, B UI morph loop, D flagship 45 s+. Append C if there's a voiceover or mascot. Fill it and savedocs/brief.md; its first line says who watches, what they believe before, and what they should believe or do after. docs/style_guide.md: fromrefs/*/analysis.mdif there's a reference, otherwise fromassets/brand.json. Palette in hex, one display face, one UI face, one accent. For "Both", define the second look asTHEMES.houseinindex.html.- Beat grid:
- Synthesized:
node music.mjs --dur <seconds> --bpm 120→audio/music.wav+audio/beats.json. - Supplied track:
python3 beats.py audio/track.wav > audio/beats.json.
- Synthesized:
- Voiceover: follow voiceover.md. Write
vo_script.json(one line per shot, ~2.3 words per second), pick the voice, generate, fix overruns. Shots then start on line times.
Step 4: Shot list gate (the only other time you stop)
Write docs/shotlist.md on the beat grid. Every shot: beat range, what's on screen (naming the
asset file), how it is made out of the previous shot (the transition), camera move, on-screen
text, voice line (if any), SFX cue. Rules: hook in the first 2 s, a change on every beat, no hold
over 1 s (2 s for explainers), one proof metric, logo + CTA at the end, at most 8 words per line
on screen.
Show the shot list compactly (and the voice name, if any) and ask for OK. Apply changes, then continue without stopping until delivery.
Step 5: Build
Edit index.html: set DUR, BPM, THEMES.brand, FONTS, ASSETS, then replace the example
SCENES (a working 11 s keynote take: wordmark squeezes into its period → dot → pill → click →
flood → 3D UI with headline → camera push → toggle → collapse into the dot → the letters spring
back out of the period, so the last frame equals the first). Scenes get absolute t; place every
event with B(n) (beat n) and let neighbouring scenes overlap so one hands off to the next.
The kit (lib/kit.js) does the Opus 5.5 grammar; use it instead of writing effects from scratch:
- Type:
maskRise(letters rise out of a line,outto leave,accentLastfor the period),words,scramble,typeOn(seeded human keystroke rhythm; put\nin the text for fixed lines;typeTimesgives the same times for key sounds),slam,countUp,strike,label,wordmark,squeeze(letters slide into the period;at: 'dot'pins the dot). - UI:
card(liftgrows the shadow with height),pill,chip,toggle,shot(screenshot or sub-rect),ring,cursorPath,cursor,ripple,glass(blurred backdrop, tint, inner edge, rim). - Camera and depth:
camPath(t, [[time, zoom, x, y], ...])(one glide per key, zoom in log space: the default camera),camera(zoom, focusX, focusY),plane3d(img, cx, cy, w, {rx, ry, rz, blur})(perspective UI plane),blurred(px, fn). - Transitions:
morphBox,states(t, table)(one morph table of beat times: size, radius, colour),flood,contract(area: trueeases by area),iris(six real blades, eased by area),portal(open a dot onto another world),goo(split / merge),mixColor. - Timing (lib/motion.js): springs
sp()for pops and landings,glide(t, 0.8)on cubic-bezier(.45,0,.15,1) for anything that travels,keyed()for multi-key glides. - Layers:
layerCtx(name)+withCtx(ctx, fn)draw a world offscreen (needslet g, notconst g). - Footage (wall shadows, sunrise, a screen recording):
node prep-footage.mjs in.mov wall→ frame sequence;FOOTAGE = {wall: 'assets/footage/wall'};footage('wall', localT, x, y, w). Keepwindow.seek = seekWith(draw)so frames load before capture. - Character:
mascot(x, y, size, t, {look, squash})(?mascot=1in the template). - World:
ambient(t, {base, pools})(soft light pools; change colours per chapter),gobo(t)(window light on a surface; warmtintfor dusk),stars(t)(night, story only). - Lens:
depthBlur(depth, focus)withblurred()for rack focus,focusIn(text, ...)(words pull into focus,outto leave). - Walkthrough:
chapterBlock(x, y, t, {kicker, light, bold, body, out})(dims, then leaves),callout(text, ax, ay, lx, ly, t)(hairline label on the UI part),laptop(...)besidephone(...),idleFloat(t)so reading holds never freeze. - The product moving itself:
stackY(t, orders, heights)(FLIP reorder / insert),typingDots,highlight(words the product understood),flyTag(a request word flies into the result field it filled),checkDot(task checkbox,pulseto invite a tap),whip(p, drawA, drawB)between chapters (the renderer adds the streak). - Entrances, exits and cuts:
arrive(t, dur)/depart(t, dur)(lib/motion.js) so no shot starts or ends still,cutBlur(t, tCut)(cut at peak blur),smear,zoomSmear(end wordmark arriving huge),matchFrame(tile, p)(full bleed ↔ tile, radius to 0).cursorPathreturnshoverandrot: pass them ascursor(x, y, press, a, { hover, rot }). - Object chains:
textRect(a word's box, to morph it into the next object),bendToRing,revealFrom(a word opening from one letter),orbit,layouts+layoutMorph(grid → ring → point),glint. - Cause → effect and data:
connector,cardSwap(status card ticking),scan,selectionBox,odometer,areaChart(draws, glowing head, scrub tooltip),pushSwap(stat replaces stat),slider,rotateWord(one slot in a fixed sentence),karaoke(voice-timed caption). - Surfaces and accents:
glass(..., { refract }),glassText(frosted title,clipto the card),gradientText,bandWipe,reflection,marquee(integrations, logos),reel,stepIndex+stepTimes(image runs on 16th notes, one click each).
Use sp(t, 'snappy'|'base'|'heavy'|'playful') for entrances and track() for anything with
several targets (cursor). Lay out in design units with DW, DH, PORTRAIT, SQUARE
so one timeline reframes per format. Ring and click on real boxes from assets/targets.json.
Sound cues live in index.html next to the events they score: the CUES array, synthesized
{t: B(4), type: 'click'} (click, pop, thump, whoosh, chime, tick, swish, riser, key) or file SFX
{t: B(4), file: 'sfx/click.wav', gain: 0.8}, placed by the file's loudest sample. Every visible
event gets a sound; typed text gets a key per letter from typeTimes. If the film has
deliberate hard cuts, list their times in window.CUTS so motion blur never smears across them.
Step 6: Critique loop (minimum 3 rounds)
node render.mjs --w 1920 --h 1080 --stills beats # one still per beat, mid-move
node render.mjs --w 1920 --h 1080 --stills beats --query theme=house # second look, if any
node render.mjs --w 1080 --h 1920 --stills 1,4,8 # spot-check each other formatOpen out/stills/contact.png with your file/image reading tool and actually look at it. Run the
"Critique pass" from briefs.md, including its slop check (with a reference, put
its sheet.png beside your contact sheet), fix the 3 worst problems, log scores in
docs/review_log.md, re-render stills. Stop only when every score is 8+ and at least 3 rounds
are done. For 45 s+ films, render an animatic first: --w 960 --h 540 --sub 1.
Step 7: Render, mix, check
Per format (16:9 1920x1080, 9:16 1080x1920, 1:1 1080x1080) and per theme:
node render.mjs --verify # same frame in any seek order
node render.mjs --w 1920 --h 1080 --out out/silent_16x9.mp4 # 60 fps, adaptive motion blur
node render.mjs --cues sfx_cues.json # window.CUES -> sfx_cues.json
node sfx.mjs sfx_cues.json out/sfx.wav
./mix.sh out/silent_16x9.mp4 out/final_16x9.mp4 audio/music.wav out/sfx.wav [--voice audio/voice.wav]
node captions.mjs --mux out/final_16x9.mp4 # only with a voiceover
./check.sh out/final_16x9.mp4 <seconds-of-fastest-action> [--loop] [--max-hold 2 for explainers]Drafts and finals stay in the workspace's out/ until they pass every check below.
--verify must pass before the first full render; a mismatch means state is carried between
frames (a cache, a counter, an unseeded random). Motion blur is --sub auto by default: each frame
block-matches the two ends of a 180° shutter (--shutter 0.5) and spaces samples so nothing moves
more than 3 px between them, up to --max-sub 32. Still frames stay single-sample and sharp; fast
moves get as many samples as they need, so there are no ghost copies. --sub N forces N samples
(--sub 1 for animatics). After a fix, re-render only the changed seconds with --from 12 --to 16.
Grab the fastest frame from the final MP4 at full resolution (ffmpeg -ss <t> -i final.mp4
-frames:v 1 out/f.png) and look at it 1:1: edges should streak, never step.
After the full render, if you can start a separate agent or subagent, have a fresh one that didn't
build the film run the critique pass on out/check/contact.png (briefs.md says how). Fix what it finds.
check.sh also runs motion-audit.mjs, which prints motion per second and lists every hold
over 1 s, every single-frame pop, dead seconds, more than 3 flashes in a second (a seizure
risk; slow the floods or cuts), (with --loop) the seam, and (when
sfx_cues.json exists) sound sync: visible events with no sound within 0.12 s, and sounds with
nothing moving. Fix every hold, every pop that isn't a deliberate downbeat cut, and every sync miss. Then open out/check/contact.png,
strip.png and phone.png and run the critique pass on the real render; fix and re-render
anything below 8. With a voiceover, listen-check by reading
audio/vo_timing.json: every line must start on its shot and none may overrun.
Confirm loudness: ffmpeg -i final.mp4 -af ebur128 -f null - 2>&1 | tail -12 shows about -14 LUFS.
Step 8: Deliver and report cost
Copy only the finished MP4s into ROOT, named <product>-<format>.mp4 (acme-16x9.mp4,
acme-9x16.mp4, acme-house-16x9.mp4 for a second look), plus <product>.srt when there's a
voiceover. If a file with that name is already there from an earlier version, add -v2, -v3
rather than overwrite it.
cp out/final_16x9.mp4 "<ROOT>/<product>-16x9.mp4"
cp out/captions.srt "<ROOT>/<product>.srt" # only with a voiceover
node cost-report.mjs --video out/final_16x9.mp4It reads exact token usage from Claude Code or Codex session logs for the user's folder (newest
session; --all-sessions for everything), adds the voice API cost from audio/vo_usage.json,
prices it with pricing.json, and writes out/cost_report.md. If no log is found (Cursor and
other IDEs don't expose usage to the agent), ask the user for the numbers on their usage page
and re-run with --model claude-opus-5-5 --input N --output N --cache-read N --cache-write N.
Only if they can't, use --transcript <chat.jsonl> and call the result a lower-bound estimate.
Then free the space: node workspace.mjs clean <product> deletes drafts, renders and
references, and keeps the film's source and the review images (a few MB) so follow-up edits
("change the CTA", "make a 9:16") are a one-line change plus re-render.
Report in this order:
- The MP4 paths in ROOT; embed
out/check/poster.pngandout/check/contact.pnginline. - One line, filled from the report: "This N-second video used X tokens and cost about $Y (Z per second of video), plus $V for the voiceover." Mention if it's an estimate, and that subscription plans bill differently from API rates.
- Final critique scores and the 2-3 things you'd improve next. Name any check you couldn't run (no voice key, no device, a step skipped) rather than leaving it out.
- One line: the source is kept in
<WORK>for edits;node SKILL_DIR/studio/workspace.mjs clean <product> --allremoves it, andworkspace.mjs listshows every workspace.
Hard rules
- The user's folder receives only the final MP4s (and the
.srtwith a voiceover). Everything else lives in the workspace. - Real product UI only. Never invent screens, features, metrics or testimonials.
- No Math.random, timers, CSS transitions or carried state in render mode.
- Banned: crossfades, whole-frame blur-ins (lens focus pulls are fine), glows, decorative particles, 3D flips, dead holds over 1 s (living reading holds up to 1.8 s, 2 s in explainers), centered title on gradient, everything fading in, corner labels/frame borders, numbered HUD chapters outside a kinetic showreel, flat voids, unreadable full-page UI, more than 3 flashes in a second.
- No film without a story device: a motif that opens it and returns, and a payoff line.
- References give the grammar, never the content, logos or characters. Don't commit downloaded reference videos; link to the creator's post.
- API keys and passwords only in
.env(~/.config/product-demo-video/.envor the user's folder); refer to them by variable name; never print them or copy them into the workspace. - Logged-in apps are read-only: navigate and screenshot, never change data.
- The product's code is read only with the user's yes, and only read: never run, build, install,
edit or commit in it, never open
.envor key files, never put its source in the film. - Native apps: the user or a deep link navigates; never tap or type in their simulator or phone.
Additional resources
- The Opus 5.5 style research (sources, measured numbers, grammar): research.md
- Brief templates and the critique prompt: briefs.md
- Voiceover providers, voices and script timing: voiceover.md
- Studio files copied into each workspace by
workspace.mjs:studio/(render.mjs, lib/motion.js, lib/kit.js, lib/browser.mjs, lib/dom-extract.js, lib/gallery.mjs, capture-site/app/dom/parts/mobile.mjs, scan-code.mjs, ref.mjs, gallery.mjs, prep-footage.mjs, motion-audit.mjs, doctor.mjs, voiceover.mjs, captions.mjs, music.mjs, sfx.mjs, beats.py, mix.sh, check.sh, cost-report.mjs, pricing.json, house-rules.md, index.html, env.example) - Method source: How to build motion design studio with Opus 5.5
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

