motion-design-edit
Skill4 starsChanged 4 months agoMITView on GitHub
Render a faceless short-form motion-graphics reel — cream paper grid, punch-color cards, taped proof cards — from a voiceover recording. No camera, no talking head; every beat is a `pr_card` / `pr_punch` / `pr_proof` graphic timed to the VO. Trigger when the user hands over a voiceover/audio file (or script + intent to record one) and wants it turned into a "paper-reel" style video, references the paper-reel aesthetic, or asks to edit a faceless/no-camera short.
In this repository5 other files it gives its agents.
- CLAUDE.mdCLAUDE.md
- make-videoSkill
- reel-scriptSkill
- video-edit-longformSkill
- video-editSkill
The file
Files
/motion-design-edit
Turns a voiceover recording into a rendered faceless reel in the "paper-reel" style: cream grid background, a hand-drawn connector line, solid punch-color interrupt cards, taped index-card proof shots. Read knowledge/paper_style.md first — it has the full design system (colors, the three graphic kinds, the 5-part story structure) and is the reference for every beat you author.
This is a sibling of the video-edit skill, not a variant of it. video-edit assumes a visible talking head and is full of rules about that (speaker matte, follow-cam, "text never overlays the face"). None of that applies here — there is no camera in the shot at all. Beats cover the entire timeline, back to back, from frame 0. Don't borrow video-edit's hf_* kinds or its speaker-first rules into a plan here.
What you need from the user
A voiceover recording (any ffmpeg-readable audio: m4a, wav, mp3). If they only have a topic, use the sibling reel-script skill first to write the script, have them record VO from it, then come back here.
Pipeline
1. Build a blank base video with the VO muxed in. The render pipeline (borrowed from video-edit, which expects a video file) needs something to treat as the source — a black frame is fine since pr_* beats cover the whole frame anyway.
ffmpeg -y -f lavfi -i "color=c=black:s=1080x1920:r=30" -i "<vo.m4a>" \
-t <exact audio duration from ffprobe> -c:v libx264 -pix_fmt yuv420p -c:a aac -b:a 192k \
"<workdir>/base.mp4"Get the exact duration first with ffprobe -v error -show_entries format=duration -of default=noprint_wrappers=1 <vo.m4a> and pass it to -t explicitly — don't rely on -shortest with an infinite lavfi source, it overshoots.
2. Transcribe.
python3 scripts/transcribe.py "<workdir>/base.mp4"Prints the words.json path (under ~/.cache/video-edit/<hash>/, shared cache namespace with video-edit — that's fine, jobs are keyed by content hash). Read the transcript back and check it against what was actually said — Whisper mishears predictably (e.g. "instructions" → "injections", "Comment" → "Common" happened on the first test). Hand-patch words.json directly for any wrong word, preserving start/end timestamps, then touch <workdir>/.polished so a future re-run doesn't overwrite your fix.
3. Author broll_plan.json in the same workdir as words.json. This is the part that actually determines whether the video feels like the reference or not — read the next section before writing beats.
4. Lint, then render.
python3 scripts/lint_plan.py <workdir>/broll_plan.json <workdir>/base.mp4
bash scripts/render.sh <workdir>/base.mp4 # preview tier — always firstPreview only. Never pass QUALITY=final until the user has seen the preview and approved it — same rule as video-edit.
Beat authoring — the part that's easy to get wrong
One idea, one cut. Never two sentences on one card. The single biggest gap between a first-draft plan and something that actually feels designed: grouping the VO's natural sentences onto one card each time a topic doesn't change. If the VO has "Most people use Claude Code wrong. Here's the fix." — that is two beats, not one. A card with two stacked sentences reads as a paragraph and kills the pacing; the reference cuts to a new card on almost every sentence boundary. When in doubt, split.
Exception: the REFRAME beat (pr_punch, pr_fullbleed: false) is allowed two short contrasting clauses on one card — that's the reference's own pattern for the "twist" moment (two halves of one realization landing together). Every other beat kind: one sentence.
pr_proof needs a real image to not look broken. Without image_path it renders as a blank taped rectangle — fine as a placeholder while you don't have an asset yet, but flag it to the user before calling a render "done." Source a real screenshot/photo/logo the same way video-edit does (WebSearch + WebFetch, real assets over generated ones) before finalizing.
Use speech_anchor on every beat, not hand-guessed start_sec/end_sec — align_to_speech.py snaps beats to the actual spoken words. Set generous end_sec (through the next beat's natural start) when hand-authoring initially; alignment + close_gaps.py tighten it automatically at render time.
Full kind reference, color tokens, and the 5-part story structure (hook → credibility → reframe → numbered tips → CTA) are in knowledge/paper_style.md.
Known gaps (fix if you hit them)
build_followcam.pylooks for a.venvthat doesn't exist for this skill — it fails and falls back to a static zoom automatically. Harmless (faceless videos don't need camera-follow anyway), just don't be alarmed by the warning.- WhisperX runs from the shared
the shared .venv(not a local one — see the comment inscripts/transcribe.py). Don't duplicate that venv here; it's ~1.3GB and there's nothing skill-specific about transcription.
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.

