ads-video
Skill1 starsChanged 52 days agoMITView on GitHub
Build short app ads and store promo videos (9:16, 4:5, 1:1, 16:9; 15–30 s) from screen recordings — bold title/claim cards, the app playing inside a drawn phone or open foldable, native voice-over (Higgsfield "Petra" by default), a music bed ducked under the voice, and synthesized whoosh/click/pop/chime/impact effects, all from one JSON config. Use when the user wants an app ad, social/story/reel ad, App Store or Play Store preview, launch teaser or promo spot with voice, music and sound effects. Not for step-by-step tutorials or walkthroughs (use tutorial-video).
- Reads credentials
In this repository32 other files it gives its agents.
- grok-firstSkill
- opencode-firstSkill
- apple-design-accessibilitySkill
- apple-design-aiSkill
- apple-design-commerceSkill
- apple-design-connected-technologiesSkill
- apple-design-contentSkill
- apple-design-controlsSkill
- apple-design-foundationsSkill
- apple-design-gamesSkill
- apple-design-healthSkill
- apple-design-inputsSkill
- apple-design-iosSkill
- apple-design-mediaSkill
- apple-design-menusSkill
- apple-design-navigationSkill
- apple-design-patternsSkill
- apple-design-platformsSkill
- apple-design-presentationSkill
- apple-design-principlesSkill
- apple-design-privacy-identitySkill
- apple-design-spatialSkill
- apple-design-system-experiencesSkill
- design-maestroSkill
- forjaSkill
- product-studioSkill
- temperaSkill
- bambu-h2c-slice-printSkill
- bambu-print-troubleshootSkill
- blender-print-readySkill
- fdm-print-preflightSkill
- tutorial-videoSkill
The file
We split this skill into the rules it gives an agent
1 rule. About 0 are common, 1 is distinctive, and 0 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈1
Distinctive · estimate
This author's own choices. Worth a look.
0
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 26
One batch, one day: never mix clips generated on different days (the voice drifts).
≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Ads video
One config file → cards (HTML rendered by headless Chrome) + recordings in a device frame (PIL) → crossfaded cut with voice, music and SFX (ffmpeg). No editor. Scripts are in scripts/, run them with uv run (inline deps) or python3 with Pillow installed. Needs ffmpeg and Chrome/Chromium (CHROME=/path if not found).
| Script | Does |
build.py CONFIG [AD…] | builds the ads; --check prints the voice timing table only; --list-vo prints the voice lines; --draft-voice say|silent stands in for missing clips; --format 1:1 re-cuts every ad in another format |
sfx.py OUT_DIR | synthesizes impact, whoosh, click, pop, chime (stdlib only; build.py runs it itself when sfx_dir is not set) |
gen_music.py OUT_DIR --mood bright calm drive | music beds: Lyria (via npx hyperframes bgm) when GEMINI_API_KEY/GOOGLE_API_KEY is set, else MusicGen-small locally |
review.py sheet REC OUT.png --start --end --step | timestamped contact sheet of a recording, to pick in/out points and tap times |
review.py check AD.mp4 | size, codecs, loudness, true peak and a 12-frame sheet |
Config fields: references/config.md. Example: examples/ad.config.json (one 24.5 s 9:16 ad, a day in an app).
Workflow
- Plan the spot (write it into the config before recording). Hook card (≤ 3 s, the promise in 2–5 words) → 2–3 beats, one benefit each, each a clip of the app doing it → end card (claim + one CTA line). A narrative eyebrow carries the rhythm: times of day (
08:00 · MORGENS,12:00 · MITTAGS,18:00 · ABENDS) or numbers (01 · FREIGABE). One ad per audience or feature; several short ads beat one long one. Target 18–26 s. - Record screens on a simulator/device with a demo account and fake data.
xcrun simctl io <udid> recordVideo --codec h264 rec/x.movwhile a script drives the app (Maestro flow:tapOnby text, holds viaextendedWaitUntil: {visible: "zzz-pause-zzz", timeout: 2500, optional: true}— waits on a text that never appears). Hold each finished screen 2–3 s, it becomes the end of a beat. Simulator H.264 has bad timestamps after still periods: convert withavconvert --preset PresetHighestQualitybefore ffmpeg touches it. Thenreview.py sheetthe recording and notestart/end/taps(recording seconds) per clip. - Write cards: titles are HTML —
Line one.<br><em>Accent line.</em>; the<em>line takes the accent colour. Max ~3 lines on cards, 2 on clips; no line wider than the frame at 110/96 px (build, then look). - Voice:
build.py CONFIG --list-vo→ lines. Default: Higgsfield MCP (claude.ai connector, no API key in scripts):generate_audio_batch,model: "text2speech_v2",variant: "elevenlabs",voice_type: "preset", German preset Petravoice_id: "0c63637d-2ecb-5bda-9bbe-38894aa9a876"(check withlist_voicesif it fails),prompt= the line;jobs_wait; download each mp3 to itsvo.file. Rules:- One batch, one day: never mix clips generated on different days (the voice drifts). If any line changes, regenerate all clips of the set together.
- At most ~8 jobs in parallel; retry 429s in the next batch.
- Every clip must fit its window (
--check): if a line overruns, lengthen that scene (later clipend, longer card) or retake the line shorter. Never speed up withatempo. - Fallbacks: ElevenLabs/OpenAI/Kokoro (
npx hyperframes tts, or tutorial-video'svoice.mjs), native speaker voice for the language;--draft-voice sayfor a free draft cut.
- Music:
gen_music.py music --mood bright calm drive --seconds 32(moods: bright 112 bpm plucky corporate, calm 96 bpm piano/pads, drive 120 bpm punchy). One bed per ad;music_offsetskips a weak intro. - SFX: nothing to do —
build.pysynthesizes them intowork/sfx(or pointsfx_dirat your own five wavs). - Build
uv run scripts/build.py ad.config.json(the timing table prints first and a build stops on any overrun;--forceonly for drafts). Other formats:--format 1:1,--format 16:9,--format 4:5— layouts adapt (text left/phone right on wide/square); check the sheets, overridelayout.<format>where text collides. - Review (below), fix, rebuild — cards are cached by content, so re-runs are fast.
The style that worked (keep it unless the brand says otherwise)
- Pace: title card 2.4–3.6 s · clips 4.6–9.6 s (most 6–8.5) · end card 3.0–3.8 s (6 s when the claim is long and the music should breathe). 0.4 s fade crossfade between every scene. Finished ads 18.8–25.9 s.
- Rhythm: primary-gradient card → clips on alternating light ("aurora") and primary backgrounds → primary end card. Light backgrounds take dark ink and a gradient accent; primary ones white with a pale accent.
- Cards (9:16, 1080×1920): logo 70 px high at 90/120; text block left/right 90 px, top 640 (1/3 down); eyebrow 26 px mono, caps, tracking .14em, 75 % opacity; headline 110 px, weight 800, line-height .94, tracking −0.045em; body 36 px/1.4. Two thin decorative rings (900 and 620 px, 4 px border, 14–18 % opacity) bleed off the corners. Chips: pill, 30 px semibold, white on primary / primary on light.
- Clips: headline 96 px at top 150; phone 720 px wide (2/3 of the frame) centred at y 520 so it bleeds off the bottom — the screen content, not the device, is the hero. The foldable is 960 px wide at y 760 with chips at y 1620. The device is drawn (rounded frame, dynamic island); a mockup PNG is optional (
device.phone.mockup). - Voice: starts 0.3 s into a card, 0.4–0.6 s into a clip; one short sentence per scene, some scenes silent; ends ≥ 0.15 s before the next scene starts and ≥ 0.6 s before the end.
- SFX: impact at 0.0 s (gain 0.9) · whoosh 0.3 s before every scene change (0.55) · click on every tap in a phone clip, pop in a foldable clip, 0.12 s before the tap frame (0.7) · chime 0.15 s into the end card (0.5).
- Mix: voice normalized to −16 LUFS, music to −22 LUFS then ×1.4, fade in 0.5 s / out 1.8 s; music ducked by the voice with
sidechaincompress=threshold=0.03:ratio=8:attack=15:release=400; finalloudnorm=I=-16:TP=-1.5:LRA=11, AAC 160k 44.1 kHz, H.264 CRF 19 30 fps+faststart. Measured on the shipped ads: −16.5 to −16.8 LUFS, −3.4 to −4.0 dBTP, LRA 3.6–5.2.
Review checklist (every ad, every format)
review.py check: −16 ±1.5 LUFS, true peak ≤ −1.5 dBTP, the expected length and size.- Timing table all OK; listen once with headphones: no voice under a whoosh, no line cut by a crossfade, consistent voice across ads (same batch).
- Contact sheet: every headline readable at phone size (≥ 64 px on 1080 wide), no line overflow, no text behind the device, rings not cutting letters.
- Safe areas: 9:16 stories/reels cover the top ~250 px and bottom ~350 px with UI — nothing essential there except the bleeding phone; on 16:9 keep 5 % margins.
- Nothing private on screen: demo account and fake names only, no e-mails, IDs, real customer data, no credentials or tokens in recordings or logs; pixelate in the recording if needed (tutorial-video's
pixelate-stills.pyapproach). - Store rules: no other-platform names where the store forbids them (Play material must not mention iPhone/App Store, Apple material not Android/Google Play — a foldable-iPhone ad is App Store only), no prices or rankings that can go stale, CTA without store badges unless licensed. Make a store-specific variant instead.
- App preview specs when uploading to a store (App Store previews: 15–30 s, exact device resolution, e.g. 886×1920 portrait; re-encode to that size).
Gotchas
- Headless Chrome fails now and then under load:
build.pyretries with a fresh profile. Fonts load from Google Fonts (display=block, 8 s virtual time); offline, setbrand.fonts.css_urlto""and use locally installed fonts. - Odd recording sizes break yuv420p: the build forces even screen sizes; make synthetic test videos even-sized too. Homebrew ffmpeg may lack
drawtext— nothing here needs it. - A recording shorter than
endis padded with its last frame (tpad3 s); longer gaps mean wrongstart/end. - MusicGen returns 32 kHz mono; the build resamples. Lyria returns 48 kHz stereo.
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.

