Demo videos from real product footage
A demo cut from real footage proves the product works. Generated imagery cannot. This skill takes a product
from "it runs in a test harness" to a 20 to 30 second video that is ready to post, and back again
when the product changes.
Start here for any product video. Use this skill first for every launch, demo or announcement video,
whatever the product. When a shot needs generated imagery (AI stills, image-to-video, a cinematic scene),
take the technique from launch-video-generation. For the post that carries the video, use
social-launch-post.
Series brief
Every video in a set follows these rules. When several agents make a set, copy them into one shared brief
and add only the per-video facts: product, style, install line, avatar look.
- Length and grid. 20 to 26 s. Pure graphic shots last 1 to 1.5 s; text shots hold for the reading pace
(see the polish rules). Cut on beats detected with
npx hyperframes beats, and skip beats as needed.
- Arc. Hook in outcome language → value (name and promise) by beat two → the demo on the musical drop →
proof from measured numbers only (the product's own tests or artifacts; never invent a number) → end card
with the exact install command and the repository URL.
- Frame 0 is a hook. The first frame is a finished graphic with the hook line, never a logo, a title or a
fade-in. A viewer decides in that frame.
- One idea per screen. One hero (text or a graphic), one visual action.
- Detailed art, not flat shapes. Draw high-quality, detailed SVG art. Flat circles and rectangles read as
childish. For isometric objects and figures, use the
anatomy skill (wheresryan22/anatomy).
- Reuse a proven style. Pick a style from the library and reuse its music prompt and
SFX pack. Inside one set, give each video a different style. Invent a new style only when the library runs
out, and write its spec after it ships.
- Names. Do not use a host platform's trademark as the product name; "for Firefox" or "runs in Chrome" is
fine. No "demo" in an on-screen product name. Read the package name from
package.json; it can differ from
the repository name.
- Plain language. Write for a viewer who knows nothing about the field: short sentences, active voice,
common words. Run the
humanizer skill on the copy.
- Music. Two instrumental variants per video, chosen by the per-second loudness curve (stage 4). SFX on
cuts, clicks and slams.
- Spend cap. Set a hard cap per video before the first call, for example $0.30 for music and SFX and
$0.15 for images. Check the provider balance before each call, log each charge to
spend.jsonl, and stop
and report when the shared balance drops under $0.50.
- Keys. Source each API key from its secrets file inside the command. Never print a key.
- Avatar. Only on the videos that get one (see the presenter section). Duck the music to about 0.26 under
the voice. Build everything else first and add the avatar last.
- Privacy is a stop-ship. Run the safe-end guard and the frame audit (stage 1 and stage 7) on every
render.
- Scope. Work in a media folder outside the product repository. Do not commit media, push, open PRs or
post anything as part of making the video.
- Report per video. Path, style, length, music model and cost, avatar look, audit result, blockers.
Stages and gates
Each stage ends at a gate. Do not start the next stage before the gate passes.
| Stage |
Output |
Gate |
| 1. Capture |
Raw footage per scenario, with the time of each moment that matters |
Every frame shows only the product window |
| 2. Plain cut |
One captioned MP4 under 60 s |
Ready to share as is; the fallback if stage 5 slips |
| 3. Storyboard |
Hook, beats, a frame table with "why" per frame |
Each claim traces to footage or a measured run |
| 4. Music |
Instrumental bed and a bar grid |
No vocals; strong first downbeat; clean last bar |
| 5. Build |
HyperFrames project per video |
hyperframes check passes with no findings |
| 6. Stills |
One still per scene, on a contact sheet |
Director notes resolved (see the checklist) |
| 7. Render |
Final MP4, verified |
Duration, streams, loudness and frames checked |
| 8. Save |
Corrections written back into this skill |
The next run starts from the corrected skill |
In collaborative mode the human approves stages 3, 6 and 7. In autonomous mode, post the storyboard
and contact sheet as heads-ups, and still do the stage 6 review yourself before you render.
1. Capture real footage
Drive the product from the same end-to-end harness that tests it, with a visible (headed) browser.
Record the screen while the harness runs. This gives footage of the real product on real sites, and the
run doubles as a test.
Record the window region, not the screen
- On macOS,
screencapture -v writes no file when a parent process stops it with a signal. A wrapper
that sends SIGINT gets nothing back. Use ffmpeg's avfoundation screen input instead, and finish it by
writing q to its stdin. That closes the file correctly.
- Park the browser window at a fixed position and size before recording (for example with
peekaboo window move/resize), then crop the capture to that region. On a Retina display the crop is in pixels, not
points: multiply the window rectangle by 2.
- Check a frame from the first test capture before the real run. The crop must show the window and nothing else.
Privacy
A full-screen capture leaks the menu bar, OS notifications and other apps' windows.
Crop to the product window. If something still covers it (a persistent notification over a side panel),
compose the video from two crops instead of touching the user's desktop.
Cut every segment before the browser closes. Harness logs are often buffered. The time a "done" line
prints can be later than the moment the window closed, and the frames between show the desktop. Sample
frames around the end of each run and cut before the last product frame.
Review the tail of every segment frame by frame before you share anything.
Cut harness segments that show a user name or account, not only the frames after the last product frame. A
harness log line such as "user":"<name>" on screen is a leak too.
Record the last product frame of each source file (for example 38.45 s), and make the build refuse any
clip whose data-media-start + data-duration × rate reads past it. A cut made 0.3 s too late put a private
notes window into a finished social video. The guard turns that mistake into a build error:
SAFE = {"flights": 38.45, "maps": 38.65} # last product frame of each source, in seconds
for cid, src, start, dur, media_start, rate in clips:
end = media_start + dur * rate
assert end <= SAFE[src], f"{cid} reads {src} to {end:.2f}s, past the safe end {SAFE[src]}s"
Scrub the final frames for any personal window, notification, email address, file path with a user name,
or token. One leak stops the release.
Third-party sites
- Use logged-out sessions. Do not record personal accounts.
- Do not send content to a third party on the user's behalf just to film it. End the scene with Cancel and
show the full send flow on a local test page.
- Probe first. Load each candidate site, check that the element under test exists (a composer, a consent
banner) and that no bot wall appears, then choose the sites to film.
- Real sites find real bugs. On one run, filming a consent add-on on a real site showed it following a policy
link out of the dialog, and saving consent with every optional switch still on. Fix the product first,
then re-record. Do not film around a bug.
2. Make a plain cut first
Cut the raw footage into one MP4 under 60 seconds: trim each scenario to its active window, add one caption
per scene, scale to 1600 px wide. This is shareable on the same day, and it is the fallback when the designed
cut slips.
- Homebrew ffmpeg builds often lack the
drawtext filter. Render each caption as a transparent PNG (any
image library) and overlay it.
- Keep scenes at real speed in the plain cut. It is the honest version.
3. Storyboard the designed cut
Write the plan before any HTML. A frame change here costs a minute; after a render it costs a cycle.
For the script itself (beats, hooks, second-by-second turns, the call to action), use
launch-video-screenwriting first, then lay out the frames below.
- Hook in outcome language. The first beat says what the viewer gains or avoids, not what the product is.
"You're one Enter away from pasting your email into an AI" beats "PII Guard detects PII".
- Value by beat two. The product name and its one-line promise land in the second scene.
- Evidence after value. Footage scenes show the promise working. Each scene gets a short topline that
says what the viewer is seeing, and changes when the footage changes.
- Proof from measured runs only. Use the numbers your harness measured (a timing, a pass count, "0
network requests"). Do not invent a statistic for impact.
- End card. Name, one-line description, where to get it. Do not claim a store listing that is still
under review. Copy the install line exactly from the README (
npm create <name>@latest,
npm i -g <name>, npm i <name>-agent). A generic npm i <name> is wrong for most packages.
- Present it as a proposal: one line saying "this video tells [audience] that [message]", then a table of
frame, beat, on screen and why. Cut any frame whose "why" does not trace back to the message.
Start from the style library. It holds every proven look, with a
table to choose from, and the presenter module, each specified down to the copy, eases, music prompt and SFX cues.
Give each video its own style. A first pass where three videos shared one look (dark radial gradient,
every line fading up, 13 s scenes) came back as "too generic". What fixed it:
- One style per video, each with its own ground, type pair, ease family and tempo band. For example: a
light gray stage with white cards and exponential whips; cream paper with a dot grid and giant serif beats;
flat color grounds that cut on the beat, with condensed display type.
- Shorter: 20 to 26 seconds, cut on beats. Fast cuts suit pure graphic frames; any frame with text holds
long enough to read (see the polish rules below).
- A clear arc: hook, stakes, value, demo on the musical drop, payoff, end card.
- Varied text entrances: masked per-word rise, velocity-matched waterfall (the first word travels furthest),
blur-to-sharp scale slam, frame-stepped typing, hard-cut word swaps on beats, wide-to-tight tracking.
- Designed seams: whip off with blur, zoom-through, mask open, color cut. Not crossfades.
Mine a real library for these patterns before writing any. HeyGen's hyperframes-launches repository has
the source of its launch films. Clone it without large media (--filter=blob:limit=300k; the media is in
Git LFS) and have a subagent extract a lookbook with file and line citations: grounds, text techniques,
seams, pacing, sound cues and story arcs.
Reference beats description. Name one or two reference videos whose pacing and type you want to match
(launch-video galleries such as whatships.com). Without a reference, an agent falls back to the same default
look: centered text, gradient background, everything fading in. For a big launch, ask for three storyboard
variants and pick one.
Polish rules
A second round on a 19-video set fixed what the owner still disliked: too much text, too many screenshots,
and cuts too fast to read. Apply these from the first storyboard.
- Reading pace. Hold a frame with a sentence until a slow reader finishes: about 3.5 s, plus 0.3 s
for each word over six. A text frame has at most about 8 words. When in doubt, cut words, not time. That gives about 8 to 10 shots in 20 to 26 s. Cut on beats, but
skip beats.
- One idea per frame. Text or a graphic is the hero, not both. One visual action per shot: never stack
a screenshot, a scan, a callout and a label in one frame.
- Graphics are the story. Draw the UI as animated SVG in the video's style: a cursor clicks, a toggle
flips, a form fills itself, a stamp lands, a shield blocks an arrow. Use at most one real screenshot per
video, as a quick "this is real" beat, cropped tight. No terminal, code, JSON or log shots unless the
viewer needs that exact command.
- Less text. At most about 40% of shots carry a text line. The rest carry a 1 to 3 word label at most.
No two shots in a row use the same entrance.
- A copy frame per video. Pick PAS (problem, agitate, solve), BAB (before, after, bridge) or AIDA, and
vary it across the set. Open on an everyday situation or the worst case, in plain words for a viewer who
knows nothing about the field. "Do you remember what you ordered yesterday? Neither does your AI."
- Proof that lands. Do not put "N/N checks pass" in every video. Prefer one concrete outcome ("0
requests left your laptop"), and keep numbers only where they carry the story.
- Frame 0 is the poster. The first frame is a finished graphic with the hook line, never blank, black
or a fade-in. Export it as
<name>-poster.png next to the MP4; it is the social thumbnail.
- Truth. Never present a scripted or test agent as a real one; label it "scripted" or "test agent".
Show no invented URL or domain, only the real repository and package names. Do not invent incidents,
and name a company as breached only when it is public record.
The why behind the polish rules
A third round of owner notes came with a reason for each. Apply the reason, not only the rule.
- One solid ground per video. A busy or shifting background competes with the graphic, so the eye
cannot find the one idea. Let the graphic carry the contrast. Use no background you would not pick on
purpose; if in doubt, use a solid or a subtle gradient in the palette.
- Graphics explain, they do not decorate. A graphic that does not show the mechanism is noise. Each
one shows cause and effect at a glance: what goes in, what is blocked, what comes out.
- Show, don't tell. People remember a story and a metaphor, not a claim. Find the product's metaphor
and build the video on it: a needle in a haystack, a web page seen as a painting, footprints that leave
damage.
- Art direction can follow the music. Under an artistic score, a literal UI demo feels flat. Lean into
the art reference: a website painted as a famous modern painting, a magnifier that sweeps a haystack of
UI and zooms in on the one target.
- Hidden versus visible must be obvious. A prompt-injection video fails if the viewer cannot see what
a person sees and what the agent reads. Show the visible page, then reveal the hidden layer with patterns
real sites use: a hidden div, an
aria-label or alt text, white-on-white or 1 px text, off-screen
text, an HTML comment, zero-width characters. Never name a real site as compromised.
- Real logos for real services. A known logo makes a viewer relate in a fraction of a second; text
cannot. Use the official SVG (Simple Icons, CC0, or the brand's press kit), unmodified and at a normal
size, only to name a service the product works with or compares against. Never imply an endorsement.
- Draw AI the way people see it now. A cartoon robot reads as dated. Use the current assistant
language: a soft orb, pulsing dots, a sparkle, a minimal face. Draw your own; do not copy a company's mark.
- The poster is a thumbnail. Header and footer text make it unreadable at small size. Use one hook
line and one strong graphic or logo.
- Exact alignment. An overlay that sits a few pixels off (a replacement word not directly over the
word it replaces) reads as fake and breaks the trust the video builds. Check every overlay in the stills
review.
4. Music bed
- Generate an instrumental per video at the exact length you need. On WaveSpeed,
elevenlabs/music takes
music_length_ms and force_instrumental. It bills $0.01 per second, so a 25 s track costs $0.25. State a tempo, "starts on a strong
downbeat, no long intro" and "clean ending on the final bar" in the prompt.
- Check every file. Equal byte sizes are normal for constant-bitrate output; compare checksums to confirm the
tracks differ.
- Detect beats (HyperFrames:
npx hyperframes beats in the project). A detector can report double the
tempo you asked for, because it counts eighth notes. Take the median gap and derive the bar: the bar at
112 BPM is 2.143 s.
- Match the genre to the story, and make the genres differ across a set of videos: a thriller cue with a
drop for a "caught in time" story, funk-pop for a "clearing clutter" story, a sparse keynote bed under a voice.
- Generate two variants per video from two models, then choose by energy curve, not by name. Under a
$0.30 cap per video, make the second variant with
minimax/music-02 (about $0.03), not
minimax/music-2.6. music-02 has no instrumental flag: put "instrumental only, no vocals" in the
prompt and run speech-to-text on the result to check for lyrics. Print the
loudness per second (ebur128) and look for the shape the story needs: a strong first second, a drop
where the demo starts, a clean end. Trim a long track to its best section instead of regenerating.
- Add SFX on the story's actions: a whoosh on each whip, a click and pop on taps, a stamp or hit on the
slams, typing under typed text, a chime on success.
elevenlabs/sound-effects-v2 on WaveSpeed costs
about $0.002 per effect. Give each SFX <audio> its own track and a slot equal to the file length.
- Put every scene boundary on a bar. Hard cuts inside a scene go on beats.
- Under a voice, duck the bed to about 0.26 while the voice plays and bring it up as the voice ends. Keep
SFX at 0.3 or less under the voice and 0.6 or less elsewhere. The voice must sit clearly on top.
- Run a final
loudnorm pass on the rendered file (about −16 to −14 LUFS integrated, peak below −1 dB).
The renderer can ignore the mix volumes, so measure the output, not the composition.
- Fade the bed in over the first 0.4 s and out over the last 1.5 s, using a volume automation lane, not a
timeline tween. Target about −16 dB mean and a peak below −1 dB.
5. Build with HyperFrames
HyperFrames renders video from HTML: data-* attributes set the timing, one paused GSAP timeline per
composition drives the animation, and the CLI checks, snapshots and renders.
- Install its skills (
npx hyperframes skills update, then the workflow skill), and start every build at
the /hyperframes router. A remix of existing footage routes to general-video.
- Write
BRIEF.md (message, destination, aspect, length, assets with source timestamps) and a STORYBOARD.md
with one block per scene, even when no human reviews it. On resume, these files are the only record of the decisions.
- When every scene is generated (stills and clips), keep the record in one
board.json that follows
video.schema.json (the storyboard-video tool in itsryanarmani/scripts) instead of the two Markdown files. See
"Keep the record in board.json" in launch-video-generation. Cuts with any real footage keep BRIEF.md and
STORYBOARD.md until the schema has a footage shot type with playback rate and safe-end limits.
- Use one project per video, with the same scene set: hook, value, footage with a topline, proof, end card.
Generate the scene files from one spec per video so all videos share one structure.
Structure rules from real lint and render failures
| Trap |
Rule |
A timed <video> inside a timed wrapper shows wrong frames, then vanishes |
Time the video. Keep its card wrapper untimed. |
| The untimed card wrapper shows as an empty frame during the hook |
Hide it in CSS (opacity: 0), then reveal it on the timeline. Not tl.set at 0. |
A fromTo on a hard-cut card renders its "from" state at build time and shows the card early |
Hard cuts use tl.set at the cut time. Zoom tweens take immediateRender: false. |
| Nested markup in a timed scene gets the warning "nested structure needs sub-composition" |
Every text scene is a sub-composition with its own timeline in scene-local time. |
| Footage zooms set from a sub-composition do nothing |
Animate the footage on the main timeline, at global time. |
| Repeated kicker elements with no tween overlap their neighbours |
One element per distinct label. |
| Decorative ghost text fails the contrast audit |
Use non-text decoration (glow, grid), or none. |
Negative z-index highlight disappears |
Order it in the DOM: the highlight first, the word in a positioned span after it. |
| Unbundled fonts warn and can fail in cloud renders |
Use a bundled pair, for example Montserrat for display and JetBrains Mono for labels. |
A presenter in the corner
For creator-style edits around a presenter (kinetic hook, gravity drop, text behind the speaker, screen push,
crash zooms, prompt bar, icon strike, giant captions, title cards), use edit-effects.
For a handheld selfie presenter instead of a studio look, see the selfie section of presenter.
A talking presenter in a corner card makes a demo personal. The script does not need to narrate the
footage; a short founder intro over it works.
- HeyGen's current API is
POST /v3/videos (type: avatar, avatar_id, voice_id, script,
aspect_ratio: "1:1", resolution, optional voice_settings.speed), then poll GET /v3/videos/{id}.
List looks with GET /v3/avatars/looks (limit at most 50). The v2 endpoints return a sunset warning.
- Prefer the user's own digital twin and cloned voice. Check every claim in the script against the
product, the same as on-screen copy. Keep the script to about 40 to 50 words.
- The default Avatar IV engine gave the twin fake, veneer-like teeth. Render with
engine: {type: "avatar_v", reference_look_id: <the filmed twin's look>} instead.
- For a new outfit or angle, edit real photos of the person, not a text-to-image portrait: an image-edit
model (WaveSpeed
google/nano-banana-2/edit) with four original photos as references, then a HeyGen
photo look in the twin's group (about $1.32 each). Use a different outfit in every avatar video of a set.
- Talking-photo models on WaveSpeed were weaker. InfiniteTalk came closest and cheapest (about $0.07 per
second) but still lost to Avatar V; Kling, LongCat and Pruna were rejected.
- For voice only, HeyGen
POST /v3/voices/speech with engine orca costs about $0.004 per second and
matched the speaker at 0.954 similarity.
- Keep the avatar's own background and put it in a rounded card. A solid card needs no alpha channel;
transparent WebM needs a matting-enabled avatar.
- Keep the avatar's sound on its
<video> (data-has-audio="true"), and duck the music under it.
- Measure the avatar clip with
ffprobe before you time the scenes. A clip can come back shorter than the
plan, and a fixed timeline then holds a silent end card for 5 s. Fit the end card to the clip.
- Vary the look across a set: other outfits and camera angles of the same twin (
GET /v3/avatars/looks)
keep several avatar videos from reading as one recording. Never repeat an outfit in a set.
- Avatar clips cost about $0.05 to $0.09 per second, with a minimum per clip. Render one clip per video at
full script length (about 15 to 20 s); a short test clip still pays the minimum.
Asset paths in sub-compositions
- A sub-composition may not reference
../. Lint rejects it. Put fonts and images next to the
sub-composition files (compositions/fonts/, compositions/banners/) and reference them relatively.
- Lint checks each sub-composition alone, so each one that uses a custom font needs its own
@font-face.
- Verify font files before use: a clone without LFS gives 130-byte pointer files, not fonts. Download open
fonts (for example from Google Fonts) as woff2 instead.
Footage handling
- Trim with
data-media-start plus data-duration. Speed up dull stretches with data-playback-rate
(1.15 to 1.65 reads as real; faster looks fake).
- Footage is muted. The only sound is the music
<audio>, which needs an id, or it is silent.
- Push in on the subject. Full browser footage is unreadable on a phone. Zoom the inner wrapper to the
element that matters (the composer, the prompt, the result, the route), and pull back before the cut.
- Zoom math, with
transform-origin at the center: to center a point at fraction p of the width at scale
s, set xPercent = -(p - 0.5) × s × 100 (the same for y). Keep |xPercent| ≤ (s − 1) / 2 × 100, or
the footage edge shows as a dark strip.
Motion doctrine
Smooth long-tail settles (power3.out), never bounce. Reveal each element when its beat arrives, not all
at once. No breathing loops, no slow drift on content in the back half of a scene. One slow, finite drift
on a background glow is enough life. All motion lives in the paused timeline; no CSS animations.
6. Review the stills
Run npx hyperframes snapshot --at <times> with one time per scene, at the moment that carries the
scene. Read the contact sheet as a director. Check each item:
- No empty cards, stray frames or bleed from a scene that is not on screen.
- Text inside the footage is legible at phone size. If not, push in.
- No cropped words in toplines or in the footage after a zoom.
- No exposed footage edge after a pan.
- Each scene stays long enough to read its topline twice.
- No desktop, notification or private content in any frame.
Give notes in camera words: "push in on the prompt at 15.4 s", "hard cut on the bar at 19.32", "pull back
before the cut". "Make it better" gets random changes. Re-snapshot after each round of fixes.
7. Render and verify
npx hyperframes render --quality high --output renders/<name>.mp4. A 40-second 1080p render takes about a
minute on a laptop.
Verify the file, not the log: ffprobe duration and streams (video and audio present), a frame grid
across the whole video, and volumedetect for loudness.
Make the frame grid with this audit script, and read every tile before anyone sees the video
(python3 audit.py <video.mp4> <grid.png> [...]):
# Frame audit: one frame every STEP seconds of each video, tiled into a grid PNG.
import subprocess, sys
from io import BytesIO
from PIL import Image
STEP, W, COLS = 1.5, 320, 6
for path, out in zip(sys.argv[1::2], sys.argv[2::2]):
d = float(subprocess.run(["ffprobe", "-v", "error", "-show_entries", "format=duration",
"-of", "csv=p=0", path], capture_output=True, text=True).stdout)
imgs, t = [], 0.7
while t < d:
png = subprocess.run(["ffmpeg", "-loglevel", "error", "-ss", f"{t:.2f}", "-i", path,
"-frames:v", "1", "-vf", f"scale={W}:-1", "-f", "image2pipe",
"-vcodec", "png", "-"], capture_output=True).stdout
imgs.append(Image.open(BytesIO(png))); t += STEP
h, rows = imgs[0].height, (len(imgs) + COLS - 1) // COLS
sheet = Image.new("RGB", (COLS * W, rows * h), "black")
for i, im in enumerate(imgs):
sheet.paste(im, ((i % COLS) * W, (i // COLS) * h))
sheet.save(out)
print(out, len(imgs), "frames")
Deliver into the product's media folder, outside the repository. Never commit media to a repo whose
checks reject committed proof files. Open the folder for the human.
Films from generated footage
Use this when the product has no capture to film (a concept, a prototype, a protocol). The pipeline is the same
as the stages above, with generated imagery in place of capture:
- Stills first. One hero still, then every shot as an edit of the hero. Contact sheet. Fix stray text. No video spend
before the person has seen the sheet.
- Image to video only for the hero shots (Veo 3.1 Fast on Vertex, 4 s, audio off). Animate the rest in the edit.
- Beat grid before any timing:
npx hyperframes beats <bed.mp3> --json. Write every time in beats.
- Build from a spec with
video-film/film_build.py (scripts repo), run hyperframes check, snapshot, render, then
loudnorm to about -16 LUFS and read a frame grid.
- Add one real screenshot as a framed proof panel. Label the film "simulated" on screen.
On a work account, use the work cloud project for Veo and Lyria, not a third-party relay. Style example with
measured numbers: anime, heavy cuts.
8. Save the corrections
After the human watches the result, write each correction back into this skill: a new trap, a note the human
repeated, a gate that was missing. The next agent reads the skill, not this chat.
Formats
- 16:9 at 1920×1080 suits X, LinkedIn and YouTube. Make this cut first.
- 9:16 needs a new layout (footage card on top, topline below), not a crop of the 16:9 cut.
- Keep each cut at 20 to 30 seconds for a feed. Put the hook in the first 2 seconds.
Related skills
launch-video-generation: generation techniques this skill draws on: image-to-video, AI stills, music
and voice models.
launch-video-screenwriting: script craft for the storyboard: beats, hooks, value turns per shot.
anatomy: detailed isometric SVG figures for graphic shots.
social-launch-post: the post that carries the video.
app-screenshots: store screenshots from the same product runs.
1---2name: ship-demo-video3description: "Use for any product's launch, demo or announcement video, built from real footage of the product working: capture from its own harness, a plain captioned cut, then a designed 20-26 s HyperFrames cut on a generated instrumental bed with cuts on detected beats. Holds the series brief, polish rules, privacy audit, spend caps and a style library of proven looks with music prompts and SFX cue maps, plus HyperFrames traps from real renders. Draws on launch-video-generation for generated stills and image-to-video."4metadata:5 source: https://skillmd.com/skills/pooriaarab/ship-demo-video6---78# Demo videos from real product footage910A demo cut from real footage proves the product works. Generated imagery cannot. This skill takes a product11from "it runs in a test harness" to a 20 to 30 second video that is ready to post, and back again12when the product changes.1314**Start here for any product video.** Use this skill first for every launch, demo or announcement video,15whatever the product. When a shot needs generated imagery (AI stills, image-to-video, a cinematic scene),16take the technique from `launch-video-generation`. For the post that carries the video, use17`social-launch-post`.1819## Series brief2021Every video in a set follows these rules. When several agents make a set, copy them into one shared brief22and add only the per-video facts: product, style, install line, avatar look.2324- **Length and grid.** 20 to 26 s. Pure graphic shots last 1 to 1.5 s; text shots hold for the reading pace25 (see the polish rules). Cut on beats detected with `npx hyperframes beats`, and skip beats as needed.26- **Arc.** Hook in outcome language → value (name and promise) by beat two → the demo on the musical drop →27 proof from measured numbers only (the product's own tests or artifacts; never invent a number) → end card28 with the exact install command and the repository URL.29- **Frame 0 is a hook.** The first frame is a finished graphic with the hook line, never a logo, a title or a30 fade-in. A viewer decides in that frame.31- **One idea per screen.** One hero (text or a graphic), one visual action.32- **Detailed art, not flat shapes.** Draw high-quality, detailed SVG art. Flat circles and rectangles read as33 childish. For isometric objects and figures, use the `anatomy` skill (wheresryan22/anatomy).34- **Reuse a proven style.** Pick a style from the [library](styles/README.md) and reuse its music prompt and35 SFX pack. Inside one set, give each video a different style. Invent a new style only when the library runs36 out, and write its spec after it ships.37- **Names.** Do not use a host platform's trademark as the product name; "for Firefox" or "runs in Chrome" is38 fine. No "demo" in an on-screen product name. Read the package name from `package.json`; it can differ from39 the repository name.40- **Plain language.** Write for a viewer who knows nothing about the field: short sentences, active voice,41 common words. Run the `humanizer` skill on the copy.42- **Music.** Two instrumental variants per video, chosen by the per-second loudness curve (stage 4). SFX on43 cuts, clicks and slams.44- **Spend cap.** Set a hard cap per video before the first call, for example $0.30 for music and SFX and45 $0.15 for images. Check the provider balance before each call, log each charge to `spend.jsonl`, and stop46 and report when the shared balance drops under $0.50.47- **Keys.** Source each API key from its secrets file inside the command. Never print a key.48- **Avatar.** Only on the videos that get one (see the presenter section). Duck the music to about 0.26 under49 the voice. Build everything else first and add the avatar last.50- **Privacy is a stop-ship.** Run the safe-end guard and the frame audit (stage 1 and stage 7) on every51 render.52- **Scope.** Work in a media folder outside the product repository. Do not commit media, push, open PRs or53 post anything as part of making the video.54- **Report per video.** Path, style, length, music model and cost, avatar look, audit result, blockers.5556## Stages and gates5758Each stage ends at a gate. Do not start the next stage before the gate passes.5960| Stage | Output | Gate |61| --- | --- | --- |62| 1. Capture | Raw footage per scenario, with the time of each moment that matters | Every frame shows only the product window |63| 2. Plain cut | One captioned MP4 under 60 s | Ready to share as is; the fallback if stage 5 slips |64| 3. Storyboard | Hook, beats, a frame table with "why" per frame | Each claim traces to footage or a measured run |65| 4. Music | Instrumental bed and a bar grid | No vocals; strong first downbeat; clean last bar |66| 5. Build | HyperFrames project per video | `hyperframes check` passes with no findings |67| 6. Stills | One still per scene, on a contact sheet | Director notes resolved (see the checklist) |68| 7. Render | Final MP4, verified | Duration, streams, loudness and frames checked |69| 8. Save | Corrections written back into this skill | The next run starts from the corrected skill |7071In collaborative mode the human approves stages 3, 6 and 7. In autonomous mode, post the storyboard72and contact sheet as heads-ups, and still do the stage 6 review yourself before you render.7374## 1. Capture real footage7576Drive the product from the same end-to-end harness that tests it, with a visible (headed) browser.77Record the screen while the harness runs. This gives footage of the real product on real sites, and the78run doubles as a test.7980### Record the window region, not the screen8182- On macOS, `screencapture -v` writes **no file** when a parent process stops it with a signal. A wrapper83 that sends SIGINT gets nothing back. Use ffmpeg's `avfoundation` screen input instead, and finish it by84 writing `q` to its stdin. That closes the file correctly.85- Park the browser window at a fixed position and size before recording (for example with `peekaboo window86 move/resize`), then crop the capture to that region. On a Retina display the crop is in pixels, not87 points: multiply the window rectangle by 2.88- Check a frame from the first test capture before the real run. The crop must show the window and nothing else.8990### Privacy9192A full-screen capture leaks the menu bar, OS notifications and other apps' windows.9394- Crop to the product window. If something still covers it (a persistent notification over a side panel),95 compose the video from two crops instead of touching the user's desktop.96- **Cut every segment before the browser closes.** Harness logs are often buffered. The time a "done" line97 prints can be later than the moment the window closed, and the frames between show the desktop. Sample98 frames around the end of each run and cut before the last product frame.99- Review the tail of every segment frame by frame before you share anything.100- Cut harness segments that show a user name or account, not only the frames after the last product frame. A101 harness log line such as `"user":"<name>"` on screen is a leak too.102- Record the last product frame of each source file (for example 38.45 s), and make the build refuse any103 clip whose `data-media-start + data-duration × rate` reads past it. A cut made 0.3 s too late put a private104 notes window into a finished social video. The guard turns that mistake into a build error:105106 ```python107 SAFE = {"flights": 38.45, "maps": 38.65} # last product frame of each source, in seconds108 for cid, src, start, dur, media_start, rate in clips:109 end = media_start + dur * rate110 assert end <= SAFE[src], f"{cid} reads {src} to {end:.2f}s, past the safe end {SAFE[src]}s"111 ```112- Scrub the final frames for any personal window, notification, email address, file path with a user name,113 or token. One leak stops the release.114115### Third-party sites116117- Use logged-out sessions. Do not record personal accounts.118- Do not send content to a third party on the user's behalf just to film it. End the scene with Cancel and119 show the full send flow on a local test page.120- Probe first. Load each candidate site, check that the element under test exists (a composer, a consent121 banner) and that no bot wall appears, then choose the sites to film.122- Real sites find real bugs. On one run, filming a consent add-on on a real site showed it following a policy123 link out of the dialog, and saving consent with every optional switch still on. Fix the product first,124 then re-record. Do not film around a bug.125126## 2. Make a plain cut first127128Cut the raw footage into one MP4 under 60 seconds: trim each scenario to its active window, add one caption129per scene, scale to 1600 px wide. This is shareable on the same day, and it is the fallback when the designed130cut slips.131132- Homebrew ffmpeg builds often lack the `drawtext` filter. Render each caption as a transparent PNG (any133 image library) and `overlay` it.134- Keep scenes at real speed in the plain cut. It is the honest version.135136## 3. Storyboard the designed cut137138Write the plan before any HTML. A frame change here costs a minute; after a render it costs a cycle.139For the script itself (beats, hooks, second-by-second turns, the call to action), use140`launch-video-screenwriting` first, then lay out the frames below.141142- **Hook in outcome language.** The first beat says what the viewer gains or avoids, not what the product is.143 "You're one Enter away from pasting your email into an AI" beats "PII Guard detects PII".144- **Value by beat two.** The product name and its one-line promise land in the second scene.145- **Evidence after value.** Footage scenes show the promise working. Each scene gets a short topline that146 says what the viewer is seeing, and changes when the footage changes.147- **Proof from measured runs only.** Use the numbers your harness measured (a timing, a pass count, "0148 network requests"). Do not invent a statistic for impact.149- **End card.** Name, one-line description, where to get it. Do not claim a store listing that is still150 under review. Copy the install line exactly from the README (`npm create <name>@latest`,151 `npm i -g <name>`, `npm i <name>-agent`). A generic `npm i <name>` is wrong for most packages.152- Present it as a proposal: one line saying "this video tells [audience] that [message]", then a table of153 frame, beat, on screen and why. Cut any frame whose "why" does not trace back to the message.154155**Start from the [style library](styles/README.md).** It holds every proven look, with a156table to choose from, and the [presenter](styles/presenter.md) module, each specified down to the copy, eases, music prompt and SFX cues.157158**Give each video its own style.** A first pass where three videos shared one look (dark radial gradient,159every line fading up, 13 s scenes) came back as "too generic". What fixed it:160161- One style per video, each with its own ground, type pair, ease family and tempo band. For example: a162 light gray stage with white cards and exponential whips; cream paper with a dot grid and giant serif beats;163 flat color grounds that cut on the beat, with condensed display type.164- Shorter: 20 to 26 seconds, cut on beats. Fast cuts suit pure graphic frames; any frame with text holds165 long enough to read (see the polish rules below).166- A clear arc: hook, stakes, value, demo on the musical drop, payoff, end card.167- Varied text entrances: masked per-word rise, velocity-matched waterfall (the first word travels furthest),168 blur-to-sharp scale slam, frame-stepped typing, hard-cut word swaps on beats, wide-to-tight tracking.169- Designed seams: whip off with blur, zoom-through, mask open, color cut. Not crossfades.170171Mine a real library for these patterns before writing any. HeyGen's `hyperframes-launches` repository has172the source of its launch films. Clone it without large media (`--filter=blob:limit=300k`; the media is in173Git LFS) and have a subagent extract a lookbook with file and line citations: grounds, text techniques,174seams, pacing, sound cues and story arcs.175176**Reference beats description.** Name one or two reference videos whose pacing and type you want to match177(launch-video galleries such as whatships.com). Without a reference, an agent falls back to the same default178look: centered text, gradient background, everything fading in. For a big launch, ask for three storyboard179variants and pick one.180181### Polish rules182183A second round on a 19-video set fixed what the owner still disliked: too much text, too many screenshots,184and cuts too fast to read. Apply these from the first storyboard.185186- **Reading pace.** Hold a frame with a sentence until a slow reader finishes: about 3.5 s, plus 0.3 s187 for each word over six. A text frame has at most about 8 words. When in doubt, cut words, not time. That gives about 8 to 10 shots in 20 to 26 s. Cut on beats, but188 skip beats.189- **One idea per frame.** Text or a graphic is the hero, not both. One visual action per shot: never stack190 a screenshot, a scan, a callout and a label in one frame.191- **Graphics are the story.** Draw the UI as animated SVG in the video's style: a cursor clicks, a toggle192 flips, a form fills itself, a stamp lands, a shield blocks an arrow. Use at most one real screenshot per193 video, as a quick "this is real" beat, cropped tight. No terminal, code, JSON or log shots unless the194 viewer needs that exact command.195- **Less text.** At most about 40% of shots carry a text line. The rest carry a 1 to 3 word label at most.196 No two shots in a row use the same entrance.197- **A copy frame per video.** Pick PAS (problem, agitate, solve), BAB (before, after, bridge) or AIDA, and198 vary it across the set. Open on an everyday situation or the worst case, in plain words for a viewer who199 knows nothing about the field. "Do you remember what you ordered yesterday? Neither does your AI."200- **Proof that lands.** Do not put "N/N checks pass" in every video. Prefer one concrete outcome ("0201 requests left your laptop"), and keep numbers only where they carry the story.202- **Frame 0 is the poster.** The first frame is a finished graphic with the hook line, never blank, black203 or a fade-in. Export it as `<name>-poster.png` next to the MP4; it is the social thumbnail.204- **Truth.** Never present a scripted or test agent as a real one; label it "scripted" or "test agent".205 Show no invented URL or domain, only the real repository and package names. Do not invent incidents,206 and name a company as breached only when it is public record.207208### The why behind the polish rules209210A third round of owner notes came with a reason for each. Apply the reason, not only the rule.211212- **One solid ground per video.** A busy or shifting background competes with the graphic, so the eye213 cannot find the one idea. Let the graphic carry the contrast. Use no background you would not pick on214 purpose; if in doubt, use a solid or a subtle gradient in the palette.215- **Graphics explain, they do not decorate.** A graphic that does not show the mechanism is noise. Each216 one shows cause and effect at a glance: what goes in, what is blocked, what comes out.217- **Show, don't tell.** People remember a story and a metaphor, not a claim. Find the product's metaphor218 and build the video on it: a needle in a haystack, a web page seen as a painting, footprints that leave219 damage.220- **Art direction can follow the music.** Under an artistic score, a literal UI demo feels flat. Lean into221 the art reference: a website painted as a famous modern painting, a magnifier that sweeps a haystack of222 UI and zooms in on the one target.223- **Hidden versus visible must be obvious.** A prompt-injection video fails if the viewer cannot see what224 a person sees and what the agent reads. Show the visible page, then reveal the hidden layer with patterns225 real sites use: a hidden div, an `aria-label` or `alt` text, white-on-white or 1 px text, off-screen226 text, an HTML comment, zero-width characters. Never name a real site as compromised.227- **Real logos for real services.** A known logo makes a viewer relate in a fraction of a second; text228 cannot. Use the official SVG (Simple Icons, CC0, or the brand's press kit), unmodified and at a normal229 size, only to name a service the product works with or compares against. Never imply an endorsement.230- **Draw AI the way people see it now.** A cartoon robot reads as dated. Use the current assistant231 language: a soft orb, pulsing dots, a sparkle, a minimal face. Draw your own; do not copy a company's mark.232- **The poster is a thumbnail.** Header and footer text make it unreadable at small size. Use one hook233 line and one strong graphic or logo.234- **Exact alignment.** An overlay that sits a few pixels off (a replacement word not directly over the235 word it replaces) reads as fake and breaks the trust the video builds. Check every overlay in the stills236 review.237238## 4. Music bed239240- Generate an instrumental per video at the exact length you need. On WaveSpeed, `elevenlabs/music` takes241 `music_length_ms` and `force_instrumental`. It bills $0.01 per second, so a 25 s track costs $0.25. State a tempo, "starts on a strong242 downbeat, no long intro" and "clean ending on the final bar" in the prompt.243- Check every file. Equal byte sizes are normal for constant-bitrate output; compare checksums to confirm the244 tracks differ.245- Detect beats (HyperFrames: `npx hyperframes beats` in the project). A detector can report double the246 tempo you asked for, because it counts eighth notes. Take the median gap and derive the bar: the bar at247 112 BPM is 2.143 s.248- Match the genre to the story, and make the genres differ across a set of videos: a thriller cue with a249 drop for a "caught in time" story, funk-pop for a "clearing clutter" story, a sparse keynote bed under a voice.250- Generate two variants per video from two models, then choose by energy curve, not by name. Under a251 $0.30 cap per video, make the second variant with `minimax/music-02` (about $0.03), not252 `minimax/music-2.6`. `music-02` has no instrumental flag: put "instrumental only, no vocals" in the253 prompt and run speech-to-text on the result to check for lyrics. Print the254 loudness per second (`ebur128`) and look for the shape the story needs: a strong first second, a drop255 where the demo starts, a clean end. Trim a long track to its best section instead of regenerating.256- Add SFX on the story's actions: a whoosh on each whip, a click and pop on taps, a stamp or hit on the257 slams, typing under typed text, a chime on success. `elevenlabs/sound-effects-v2` on WaveSpeed costs258 about $0.002 per effect. Give each SFX `<audio>` its own track and a slot equal to the file length.259- Put every scene boundary on a bar. Hard cuts inside a scene go on beats.260- Under a voice, duck the bed to about 0.26 while the voice plays and bring it up as the voice ends. Keep261 SFX at 0.3 or less under the voice and 0.6 or less elsewhere. The voice must sit clearly on top.262- Run a final `loudnorm` pass on the rendered file (about −16 to −14 LUFS integrated, peak below −1 dB).263 The renderer can ignore the mix volumes, so measure the output, not the composition.264- Fade the bed in over the first 0.4 s and out over the last 1.5 s, using a volume automation lane, not a265 timeline tween. Target about −16 dB mean and a peak below −1 dB.266267## 5. Build with HyperFrames268269HyperFrames renders video from HTML: `data-*` attributes set the timing, one paused GSAP timeline per270composition drives the animation, and the CLI checks, snapshots and renders.271272- Install its skills (`npx hyperframes skills update`, then the workflow skill), and start every build at273 the `/hyperframes` router. A remix of existing footage routes to `general-video`.274- Write `BRIEF.md` (message, destination, aspect, length, assets with source timestamps) and a `STORYBOARD.md`275 with one block per scene, even when no human reviews it. On resume, these files are the only record of the decisions.276- When every scene is generated (stills and clips), keep the record in one `board.json` that follows277 `video.schema.json` (the `storyboard-video` tool in itsryanarmani/scripts) instead of the two Markdown files. See278 "Keep the record in board.json" in `launch-video-generation`. Cuts with any real footage keep `BRIEF.md` and279 `STORYBOARD.md` until the schema has a footage shot type with playback rate and safe-end limits.280- Use one project per video, with the same scene set: hook, value, footage with a topline, proof, end card.281 Generate the scene files from one spec per video so all videos share one structure.282283### Structure rules from real lint and render failures284285| Trap | Rule |286| --- | --- |287| A timed `<video>` inside a timed wrapper shows wrong frames, then vanishes | Time the video. Keep its card wrapper untimed. |288| The untimed card wrapper shows as an empty frame during the hook | Hide it in CSS (`opacity: 0`), then reveal it on the timeline. Not `tl.set` at 0. |289| A `fromTo` on a hard-cut card renders its "from" state at build time and shows the card early | Hard cuts use `tl.set` at the cut time. Zoom tweens take `immediateRender: false`. |290| Nested markup in a timed scene gets the warning "nested structure needs sub-composition" | Every text scene is a sub-composition with its own timeline in scene-local time. |291| Footage zooms set from a sub-composition do nothing | Animate the footage on the main timeline, at global time. |292| Repeated kicker elements with no tween overlap their neighbours | One element per distinct label. |293| Decorative ghost text fails the contrast audit | Use non-text decoration (glow, grid), or none. |294| Negative `z-index` highlight disappears | Order it in the DOM: the highlight first, the word in a positioned span after it. |295| Unbundled fonts warn and can fail in cloud renders | Use a bundled pair, for example Montserrat for display and JetBrains Mono for labels. |296297### A presenter in the corner298299For creator-style edits around a presenter (kinetic hook, gravity drop, text behind the speaker, screen push,300crash zooms, prompt bar, icon strike, giant captions, title cards), use [edit-effects](references/edit-effects.md).301For a handheld selfie presenter instead of a studio look, see the selfie section of [presenter](styles/presenter.md).302303A talking presenter in a corner card makes a demo personal. The script does not need to narrate the304footage; a short founder intro over it works.305306- HeyGen's current API is `POST /v3/videos` (`type: avatar`, `avatar_id`, `voice_id`, `script`,307 `aspect_ratio: "1:1"`, `resolution`, optional `voice_settings.speed`), then poll `GET /v3/videos/{id}`.308 List looks with `GET /v3/avatars/looks` (`limit` at most 50). The v2 endpoints return a sunset warning.309- Prefer the user's own digital twin and cloned voice. Check every claim in the script against the310 product, the same as on-screen copy. Keep the script to about 40 to 50 words.311- The default Avatar IV engine gave the twin fake, veneer-like teeth. Render with312 `engine: {type: "avatar_v", reference_look_id: <the filmed twin's look>}` instead.313- For a new outfit or angle, edit real photos of the person, not a text-to-image portrait: an image-edit314 model (WaveSpeed `google/nano-banana-2/edit`) with four original photos as references, then a HeyGen315 photo look in the twin's group (about $1.32 each). Use a different outfit in every avatar video of a set.316- Talking-photo models on WaveSpeed were weaker. InfiniteTalk came closest and cheapest (about $0.07 per317 second) but still lost to Avatar V; Kling, LongCat and Pruna were rejected.318- For voice only, HeyGen `POST /v3/voices/speech` with engine `orca` costs about $0.004 per second and319 matched the speaker at 0.954 similarity.320- Keep the avatar's own background and put it in a rounded card. A solid card needs no alpha channel;321 transparent WebM needs a matting-enabled avatar.322- Keep the avatar's sound on its `<video>` (`data-has-audio="true"`), and duck the music under it.323- Measure the avatar clip with `ffprobe` before you time the scenes. A clip can come back shorter than the324 plan, and a fixed timeline then holds a silent end card for 5 s. Fit the end card to the clip.325- Vary the look across a set: other outfits and camera angles of the same twin (`GET /v3/avatars/looks`)326 keep several avatar videos from reading as one recording. Never repeat an outfit in a set.327- Avatar clips cost about $0.05 to $0.09 per second, with a minimum per clip. Render one clip per video at328 full script length (about 15 to 20 s); a short test clip still pays the minimum.329330### Asset paths in sub-compositions331332- A sub-composition may not reference `../`. Lint rejects it. Put fonts and images next to the333 sub-composition files (`compositions/fonts/`, `compositions/banners/`) and reference them relatively.334- Lint checks each sub-composition alone, so each one that uses a custom font needs its own `@font-face`.335- Verify font files before use: a clone without LFS gives 130-byte pointer files, not fonts. Download open336 fonts (for example from Google Fonts) as woff2 instead.337338### Footage handling339340- Trim with `data-media-start` plus `data-duration`. Speed up dull stretches with `data-playback-rate`341 (1.15 to 1.65 reads as real; faster looks fake).342- Footage is muted. The only sound is the music `<audio>`, which needs an `id`, or it is silent.343- **Push in on the subject.** Full browser footage is unreadable on a phone. Zoom the inner wrapper to the344 element that matters (the composer, the prompt, the result, the route), and pull back before the cut.345- Zoom math, with `transform-origin` at the center: to center a point at fraction `p` of the width at scale346 `s`, set `xPercent = -(p - 0.5) × s × 100` (the same for y). Keep `|xPercent| ≤ (s − 1) / 2 × 100`, or347 the footage edge shows as a dark strip.348349### Motion doctrine350351Smooth long-tail settles (`power3.out`), never bounce. Reveal each element when its beat arrives, not all352at once. No breathing loops, no slow drift on content in the back half of a scene. One slow, finite drift353on a background glow is enough life. All motion lives in the paused timeline; no CSS animations.354355## 6. Review the stills356357Run `npx hyperframes snapshot --at <times>` with one time per scene, at the moment that carries the358scene. Read the contact sheet as a director. Check each item:359360- No empty cards, stray frames or bleed from a scene that is not on screen.361- Text inside the footage is legible at phone size. If not, push in.362- No cropped words in toplines or in the footage after a zoom.363- No exposed footage edge after a pan.364- Each scene stays long enough to read its topline twice.365- No desktop, notification or private content in any frame.366367Give notes in camera words: "push in on the prompt at 15.4 s", "hard cut on the bar at 19.32", "pull back368before the cut". "Make it better" gets random changes. Re-snapshot after each round of fixes.369370## 7. Render and verify371372- `npx hyperframes render --quality high --output renders/<name>.mp4`. A 40-second 1080p render takes about a373 minute on a laptop.374- Verify the file, not the log: `ffprobe` duration and streams (video and audio present), a frame grid375 across the whole video, and `volumedetect` for loudness.376- Make the frame grid with this audit script, and read every tile before anyone sees the video377 (`python3 audit.py <video.mp4> <grid.png> [...]`):378379 ```python380 # Frame audit: one frame every STEP seconds of each video, tiled into a grid PNG.381 import subprocess, sys382 from io import BytesIO383 from PIL import Image384385 STEP, W, COLS = 1.5, 320, 6386 for path, out in zip(sys.argv[1::2], sys.argv[2::2]):387 d = float(subprocess.run(["ffprobe", "-v", "error", "-show_entries", "format=duration",388 "-of", "csv=p=0", path], capture_output=True, text=True).stdout)389 imgs, t = [], 0.7390 while t < d:391 png = subprocess.run(["ffmpeg", "-loglevel", "error", "-ss", f"{t:.2f}", "-i", path,392 "-frames:v", "1", "-vf", f"scale={W}:-1", "-f", "image2pipe",393 "-vcodec", "png", "-"], capture_output=True).stdout394 imgs.append(Image.open(BytesIO(png))); t += STEP395 h, rows = imgs[0].height, (len(imgs) + COLS - 1) // COLS396 sheet = Image.new("RGB", (COLS * W, rows * h), "black")397 for i, im in enumerate(imgs):398 sheet.paste(im, ((i % COLS) * W, (i // COLS) * h))399 sheet.save(out)400 print(out, len(imgs), "frames")401 ```402- Deliver into the product's media folder, outside the repository. Never commit media to a repo whose403 checks reject committed proof files. Open the folder for the human.404405## Films from generated footage406407Use this when the product has no capture to film (a concept, a prototype, a protocol). The pipeline is the same408as the stages above, with generated imagery in place of capture:4094101. Stills first. One hero still, then every shot as an edit of the hero. Contact sheet. Fix stray text. No video spend411 before the person has seen the sheet.4122. Image to video only for the hero shots (Veo 3.1 Fast on Vertex, 4 s, audio off). Animate the rest in the edit.4133. Beat grid before any timing: `npx hyperframes beats <bed.mp3> --json`. Write every time in beats.4144. Build from a spec with `video-film/film_build.py` (scripts repo), run `hyperframes check`, snapshot, render, then415 `loudnorm` to about -16 LUFS and read a frame grid.4165. Add one real screenshot as a framed proof panel. Label the film "simulated" on screen.417418On a work account, use the work cloud project for Veo and Lyria, not a third-party relay. Style example with419measured numbers: [anime, heavy cuts](styles/anime-heavy-cut.md).420421## 8. Save the corrections422423After the human watches the result, write each correction back into this skill: a new trap, a note the human424repeated, a gate that was missing. The next agent reads the skill, not this chat.425426## Formats427428- 16:9 at 1920×1080 suits X, LinkedIn and YouTube. Make this cut first.429- 9:16 needs a new layout (footage card on top, topline below), not a crop of the 16:9 cut.430- Keep each cut at 20 to 30 seconds for a feed. Put the hook in the first 2 seconds.431432## Related skills433434- `launch-video-generation`: generation techniques this skill draws on: image-to-video, AI stills, music435 and voice models.436- `launch-video-screenwriting`: script craft for the storyboard: beats, hooks, value turns per shot.437- `anatomy`: detailed isometric SVG figures for graphic shots.438- `social-launch-post`: the post that carries the video.439- `app-screenshots`: store screenshots from the same product runs.