Communitygithub.com

video-edit-compare

Use when a completed cut-as-code delivery needs a strict side-by-side review against its original source, including cuts, varispeed, color grade, and graphics.

What is video-edit-compare?

video-edit-compare is a Claude Code agent skill that use when a completed cut-as-code delivery needs a strict side-by-side review against its original source, including cuts, varispeed, color grade, and graphics.

Works with~Claude Code~Codex CLI~Cursor
npx skills add https://github.com/WhiteTowerAI/cut-as-code/tree/main/skills/video-edit-compare

Ask in your favorite AI

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

Documentation

Video Edit Compare

Render the single V1 comparison mode: original-vs-final-source-time.

The left panel plays the original continuously. The right panel projects actual final-delivery pixels back onto the original source clock. Dropped ranges are black, kept final ranges are stretched to their source duration, and audio is copied from the original.

Dependencies

Require ffmpeg/ffprobe on PATH and Python with Pillow. Check them before processing media.

Inputs

  • work/timeline.json: canonical chronological source-to-program mapping
  • input/original-video.mp4: original source
  • final/final-video.mp4: actual delivery after all selected edits

Do not compare against a cut plan or an intermediate render. The final pixels are required so the review includes grade and cards.

Render

python scripts/make_compare.py TIMELINE.json SOURCE.mp4 FINAL.mp4 OUT.mp4 [--filter-only]

Project example:

python skills/video-edit-compare/scripts/make_compare.py `
  work/timeline.json `
  input/original-video.mp4 `
  final/final-video.mp4 `
  review/04-edit-compare/original-vs-final-source-time.mp4

The script always writes the durable work/edit-compare/compare-plan.json. It writes PIL label images and the generated filtergraph under disposable work/cache/. --filter-only creates the plan and inspectable cache files without rendering the video.

Mapping Rules

  • Keep source clips chronological, non-overlapping, and single-source.
  • Trim each clip's program_range from the final video.
  • Stretch it by the clip speed to recover its source_range duration.
  • Insert black for every source-time gap.
  • Map the original audio without retiming it.
  • Derive every part's start and end frame from its absolute source-time boundaries, then trim to that frame count and rebuild PTS at the timeline's rational FPS. Never round each part duration independently; cumulative concat rounding causes source-time drift.

Reordered, duplicated, overlapping, nonlinear-speed, or multi-source timelines are unsupported in V1 and must fail validation.

Verify

After rendering, the script automatically checks the complete output:

  • duration equals source_duration_s within one frame;
  • output width is twice the source width;
  • every dropped-range sample is black on the right;
  • kept-range samples match the corresponding final program frames;
  • original audio remains continuous and on the source clock.

On success it writes review/04-edit-compare/comparison-summary.md with a passing verification result. When work/project.json exists, it also adds or refreshes the original-vs-final-source-time review node, records exact active operation revisions in based_on, and refreshes START-HERE.md. Do not create a separate checks manifest.

This comparison is a review deliverable, not a default preview. Generate it after the final render exists.

Individual skills in this repo

This repo contains 8 individual skills — each has its own dedicated page.

Related Skills