video-cut

What this does

Takes an agent-authored clip plan (which source ranges to keep, in order) and:

1. Validates + enriches it into clip_plan_validated.json (clip ids, source/output times, total duration, overlap checks). Boundaries are then snapped to natural pauses (SNAP_CLIP_LINE_END) and nudged clear of the original footage's hard cuts (SCENE_CUT_SNAP, see below).
2. Concatenates those source ranges into a single edited_source.mp4.
3. (Orchestrated path — default) Stops here; the agent writes narration.json against the output timeline (0 .. total seconds). Invoked by recap.py with --no-narration-map.
4. (Legacy single-pass path) When invoked directly without --no-narration-map: maps narration written in original-video time onto the cut output timeline → narration_mapped.json.

It is stateless: given the same inputs it reproduces the same outputs. Caching is just "reuse edited_source.mp4 if it is newer than clip_plan.json".

Input contract

work_dir/clip_plan.json — a JSON array, or {"clips": [...]}. Each clip:

{"start": 12.0, "end": 28.5, "reason": "inciting incident"}

start/end are original-video seconds (aliases source_start/source_end or in/out accepted). Optionally target_duration at the object top level (e.g. "10m").

work_dir/narration.json (optional, legacy single-pass path only) — segments with original-video start/end + narration text, consumed only when --no-narration-map is NOT passed. Each segment may carry source_clip_id to disambiguate when overlapping clips are allowed.

Running the scripts below — the scripts/… paths are relative to this skill's own directory (the folder containing this SKILL.md). Claude Code runs commands from there, so they work as written. If your harness runs commands from the project root instead (opencode / Codex / OpenClaw commonly do), prefix this skill's absolute directory — e.g. <skill-dir>/scripts/…, using the directory your harness reports when it loads the skill. The scripts self-locate from their own path, so once started by the correct path they resolve their sibling skills and assets regardless of the working directory.

Run

python3 scripts/cut.py <video> --work-dir <work_dir> \
  [--target-duration 10m] [--clip-padding 0] [--allow-overlap]

Output contract

• clip_plan_validated.json — normalized clips with clip_id, source_start/end, output_start/end, duration.
• edited_source.mp4 — the concatenated source video (the new shortened timeline).
• narration_mapped.json — (legacy single-pass path only) narration with start/end rewritten to output time and source_clip_id set. NOT produced in the orchestrated flow (recap.py --no-narration-map).

In the orchestrated flow, downstream (voiceover + assemble) treat edited_source.mp4 as the video and narration.json (written by the agent against the output timeline) as the narration.

Notes

• In the legacy single-pass path, timestamps in clip_plan.json and narration.json are original-video time; this tool does the source→output remap. In the orchestrated path, narration.json is written by the agent directly in output-timeline time — no remap is performed.
• Overlapping/duplicate source ranges raise an error unless --allow-overlap is set; with overlap, narration should set source_clip_id.
• Segments that fall outside every kept clip are dropped (logged).
• Shot-change-aware boundaries (SCENE_CUT_SNAP, default on). A clip boundary that lands a few tenths of a second from a hard cut in the original footage shows a brief sliver of the adjacent shot that then cuts again — a visible 闪烁/flicker at the edit point. After the natural-pause snap, each source_start is moved forward onto, and each source_end back onto, any original shot-change within SCENE_CUT_SNAP_MARGIN (default 0.5s; detected with ffmpeg's scene metric at SCENE_CUT_DETECT_THRESHOLD, default 0.4). Boundaries already on a cut, or with no nearby cut, are untouched; snaps that would shrink a clip below ~0.5s are skipped. Set SCENE_CUT_SNAP=0 to disable.

What this skill does NOT do

• Does NOT re-transcribe or re-analyze the video.
• Does NOT write narration, and does NOT pick clips for you — it consumes an agent-authored clip_plan.json.
• Does NOT re-encode beyond the concatenation/remap needed to build the cut source.