remotion-ad-pipeline
Skill4 starsChanged 54 days agoMITView on GitHub
Build code-driven 2D motion video ads and product stills with Remotion (React), plus the voice-over, captions and music mix used by every ad. Use when the ad doesn't need full 3D, for product/store stills and thumbnails, or for generating VO (voice clone TTS plus whisper word timings) and the loudness-normalized mix. Uses kit/audio/.
- Installs packages
In this repository4 other files it gives its agents.
- ad-copywritingSkill
- ad-video-directorSkill
- blender-ad-pipelineSkill
- render-precheckSkill
The file
We split this skill into the rules it gives an agent
3 rules. About 0 are common, 2 are distinctive, and 1 is very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈2
Distinctive · estimate
This author's own choices. Worth a look.
1
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 11
From scratch, in code: read real data from the app's repo (content, pricing, feature lists) and rebuild its surfaces as animated components.
- Distinctive: line 17
Split screens: before/after halves get equal bands, with captions kept out of both.
- Very specific: line 53
Synthesized music: write samples at integer offsets.
≈ “Common” is an estimate: at least 1 in 5 of the 372,429 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
Remotion ads, stills and audio
Remotion motion ads
- Project: a standalone Remotion project for marketing. Compositions go in
src/<set>/, assets inpublic/<set>/. Render at 1080x1920 and 30fps. - From scratch, in code: read real data from the app's repo (content, pricing, feature lists) and rebuild its surfaces as animated components. Don't screenshot-pan and don't open a browser to capture the site.
- Motion language: borrow it from the product's own landing page, e.g. floating rotated cards, growing bars on
cubic-bezier(0.22,1,0.36,1). Don't reuse the site's calm editorial typography in ads. - Pacing:
- A visual change every 0.8–1.5s, and a bigger scene or palette shift every 3–4s.
- Fades of at most 6–8 frames. Hard cuts are fine.
- Hold each focal idea long enough to read.
- Split screens: before/after halves get equal bands, with captions kept out of both.
- 3D inside Remotion:
@remotion/threewith--gl=angle. - Check your work by rendering stills at key frames with safe-zone overlays before a full render (see
render-precheck).
Product stills (store pages, thumbnails)
- Design stills separately. Stills reusing the video's cards, or recoloured card walls, get rejected as "same design".
- A direction that works:
- one glossy 3D hero object per image,
- a bold two-line headline,
- outcome-led copy,
- a small "More … on the way" pill if the product is still growing.
- Leave off: a price, and content counts as the lead.
Voice-over (kit/audio/vo.py)
- Engine: Qwen3-TTS (1.7B) via
mlx-audioon Apple Silicon, cloning a voice fromaudio/voice-ref.wavand its transcript (AMV_VOICE_REF_TEXT). Only clone a voice you have rights to. - Takes: each script line is generated with 3 seeds. The best take is chosen by whisper word accuracy, pacing and pitch variation.
- Timing: the whole track is played at tempo 1.05, then mlx-whisper large-v3-turbo produces word timings.
- Output:
audio/<ad>-vo.wavandaudio/<ad>-vo.json(duration,lines,words). This JSON drives captions and animation cues in both pipelines. - Voices to avoid: stock TTS voices (e.g. Kokoro presets) get called "fake voice, AI slop". Aim for a natural, conversational read at about 170–190 wpm.
pip install mlx-audio mlx-whisper librosa soundfile
AMV_PROJECT=. AMV_VOICE_REF_TEXT="exact words in the reference clip" python kit/audio/vo.py myadscripts.json looks like {"myad": [{"id": "hook", "text": "…", "pause": 0.25}, …]}.
Mix (kit/audio/mix.sh)
BED=music/bed.wav kit/audio/mix.sh myad produces audio/myad-mix.wav. The VO is highpassed and compressed, the bed is sidechain-ducked under it, and loudnorm targets −14 LUFS with −1.5 dBTP true peak.
Music and SFX
- Music: a low bed (minor-key lo-fi or soft trap) ducked about 12dB under the VO.
- Synthesized music: write samples at integer offsets. Round the start once, then add +j. Flooring
(start+t)*SRper sample skips or duplicates samples and causes a shrill, screeching artefact. - SFX: minimal and premium. Frantic stingers on every cut read as a game ad. Retime SFX when animations move.
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.

