Communitygithub.com

video-cutaways

Plan and execute cutaways and B-roll from images, footage, screenshots and basic animation over an established A-cut. Use for visual coverage, proof, reframing or simple motion without changing story timing.

What is video-cutaways?

video-cutaways is a Claude Code agent skill that plan and execute cutaways and B-roll from images, footage, screenshots and basic animation over an established A-cut. Use for visual coverage, proof, reframing or simple motion without changing story timing.

Works with✓Claude Code~Codex CLI~Cursor
npx skills add https://github.com/georgewangyu/GeorgeVideoAgent/tree/HEAD/skills/video-cutaways

Ask in your favorite AI

Open a new chat with this agent skill pre-loaded.

Documentation

Cutaways and B-roll

Cover a story beat only when the picture adds evidence, context, action or comprehension. Read the project contract.

Plan source-backed coverage [bl: invariant]

Bind the plan to the A-cut hash, duration and timebase. Inventory supplied images, footage and screenshots first: canonical source/hash, rights, privacy, relevant visible state and source in/out. Record the story job, output interval, source range, crop/fit, motion, layer and caption/face exclusion for each placement. Distinct files/crops of one image are the same underlying asset for reuse checks. A quiet A-roll beat is a valid choice; no coverage quota is required.

Use genuine screenshots/capture for software behavior and readable proof. Illustration can explain a concept but cannot impersonate a completed login, transaction or product interaction. Missing proof is a gap, not permission to invent it. Check complete sensitive shot ranges and transition neighbors, not just a reported frame. Keep all private source references with the project.

Execute on the locked clock [bl: invariant]

Overlay or replace picture over exact output intervals without changing A-cut speech, duration or shot/story positions. Cutaway source audio defaults off unless a planned native-sound event calls for it. A requested story insertion or timing change returns to Rough cut (A-cut), records an old→new map and invalidates affected downstream layers.

For basic animation, use the available editor or a verified FFmpeg filter: position/scale movement, proportional crop, fade, simple text/shape emphasis. Choose restrained movement with readable holds. Check animation extrema and start/end frames; do not add zoom to already-moving footage automatically. Apply display rotation before contain/cover scaling; never stretch both axes. Avoid loops or frozen holds masquerading as distinct footage. No new renderer, provider upload or asset purchase is implied.

With FFmpeg, explicit trim/setpts and enable expressions can place bounded sources; overlay alpha must be correct and secondary inputs duration-bounded. Preserve the original speech bed, map streams explicitly and use one final composite from primary/lossless inputs. Tool syntax must match the installed build; test a bounded representative scene before full assembly.

Verify and hand off [bl: invariant]

Compare A-cut and composite duration/frame count and audio mapping. Decode representative first/last/motion-extreme frames and each change in visual type; watch complete cutaway intervals and transitions for truthful state, legibility, privacy, geometry, occlusion and unintended black/freeze. Recheck caption collisions after composition. Still frames do not establish smooth motion. Return editable placements, source provenance, actual render/hash and honest viewing coverage. Mark absent visual playback pending.

References

Individual skills in this repo

This repo contains 4 individual skills — each has its own dedicated page.

Related Skills