Communitygithub.com

video-rough-cut

Assemble an editable rough cut from supplied footage and an editing brief, selecting complete takes and trimming repeats, false starts, and pauses without losing meaning or natural delivery. Use for a dialogue-led A-cut or a visual story/demo assembly, before finishing and effects.

What is video-rough-cut?

video-rough-cut is a Claude Code agent skill that assemble an editable rough cut from supplied footage and an editing brief, selecting complete takes and trimming repeats, false starts, and pauses without losing meaning or natural delivery. Use for a dialogue-led A-cut or a visual story/demo assembly, before finishing and effects.

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

Ask in your favorite AI

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

Documentation

Rough cut (A-cut): dialogue/story, takes, timing

Deliver a source-linked decision list and, when execution tools and evidence permit, an inspectable rough export. Prioritize coherent story and natural speech over a target number of cuts. This skill supplies a workflow, not a renderer or a universal duplicate-removal algorithm.

Source and intent contract [bl: invariant]

Read the shared project contract. For music-led material, use video-music for an early phrase/beat map before fixing story timing. Do not force speech onto a beat grid.

Use the supplied editing brief; if material choices are unresolved, use video-editing-brief only for those gaps. Preserve originals, exact required lines, factual corrections and must-keeps. Write derived files separately. Inventory source identity, duration, streams, frame rate/timebase and transcript coverage when accessible. Distinguish media actually watched/listened from transcript-only, metadata or user description. Do not turn missing evidence into an inspected take or a precise cut point.

A dialogue-led A-cut builds the spoken story before decorative coverage. A visual story or demo builds around visible actions, cause/result and required screen states; silence can carry the essential event. Neither mode requires the other. Watch silent material before excluding it. A partial transcript cannot authorize removal from the untranscribed remainder.

If proxies or separate audio are needed, preserve relink identities and the mapping to original time. Record audio authority, offset convention and any clock drift. Verify sync with observable events across the selected span; never equate a filename match, single clap or ASR timestamp with sustained sync. Unknown sync stays unresolved; do not silently replace audio or change speed.

Take selection and natural trimming [bl: soft]

Compare candidate takes in context, including their lead-in and ending. Choose the strongest complete truthful delivery for each story job, rather than always the last, shortest, or most transcript-similar take. Preserve alternatives.

Word timing comes first: repeats can only be found in words the transcript kept. If a source has no word-timed transcript (segment or chunk timestamps are not enough), run the bundled transcriber on the full, untrimmed source. ${CLAUDE_PLUGIN_ROOT} is the installed package root in Claude Code; other agents use the package root path:

python3 "${CLAUDE_PLUGIN_ROOT}/skills/video-rough-cut/scripts/transcribe.py" SOURCE.mov -o source-words.json

It uses an already-importable backend and keeps fillers and restarts where the model allows. If none is importable it installs nothing, prints a one-off uv run --with … command, and exits 3: ask the user before any download or cloud upload. Without word timing, mark the repeat audit as not run. Check existing transcripts for silence trimming or chunk-only timing before reuse.

For dialogue with word timestamps, run the bundled trim audit once per recording or speaker track before selecting takes:

python3 "${CLAUDE_PLUGIN_ROOT}/skills/video-rough-cut/scripts/trim_audit.py" source-words.json

Supply a word list (or a JSON object with words), with text, start and end in seconds. Add --protect START END for each intentional beat and optionally --min-gap SECONDS to surface quiet intervals. A transcript gap can contain a reaction or untranscribed action. Inspect both deliveries for exact repeats; for a flagged restart, hear the intervening clause before dropping anything. No timing evidence? Use the source audio and mark uncertain edges for review.

Review the proposed story for repeated meaning as well as repeated wording: paraphrases may do the same job, while a recurring hook, callback, correction, question/answer or deliberate emphasis may do different jobs. Record why a repeat is removed or retained. Keep the best complete unit and any distinct information. For a restart or false start, compare the whole attempted thought, not just its repeated prefix.

Boundary evidence [bl: invariant]

Do not splice words across takes to invent a claim or delivery. When a correction changes a fact, retain the corrected meaning and necessary context; do not join an old number to a new qualifier.

Treat ASR, alignment, silence and similarity detections as candidates, not cut authority. Raw verbatim words can expose restarts that normalized text hides; ASR can also omit repeats. Check retained interiors and outer edges as well as joins. Do not clear repeats from a high source/render word-match score alone.

Resolve suspicious cut edges against audible speech and neighboring words. Zero-length, stretched, overlapping or missing word timings, quiet consonants and overlapping speakers make an edge uncertain. Forced alignment can refine candidate timing but does not prove the transcript or safe phoneme boundary. Keep enough source context for complete sounds, breaths and natural pauses; choose conservative phrase boundaries or leave a candidate unresolved if listening cannot establish a safe trim. Do not invent a universal millisecond handle, silence threshold, pause quota or default speed-up. A quiet breath, reaction, visual action or meaningful pause is not dead air by definition.

Editable decisions and execution [bl: invariant]

Keep original source timestamps as the authority even after several passes; maintain the separate source-to-output map. Use stable decision IDs and include kept ranges, removed ranges/reasons, competing takes, unresolved edges, and intentional repetitions. Use the handoff contract when choosing a portable table or JSON representation. Mark provisional ranges explicitly so they are not mistaken for executable cuts.

Check range bounds and ordering, timestamp units/timebases, mapped audio and video durations, and the intended timeline. Reusing a source range, overlapping shots, gaps or split audio/video edits must be intentional and documented; otherwise repair them before conform. Preserve source cadence unless the output spec requires conversion, and document mixed/VFR handling and drift risk.

Use the available editor or cut/conform tool, verifying its supported inputs. FFmpeg can execute trims, conform and render; it does not decide which take is semantically redundant or which delivery is best. Do not claim a bundled backend or silently install/upload material. If execution is unavailable, return the editable proposed decisions and specific evidence gaps, not a fabricated export. Before media work, respect active recording/jobs and resource constraints; start with bounded input-seek samples, serialize heavy decodes, write new outputs and stop only task-owned processes. Do not launch an unbounded full render to test an uncertain join. Keyframe-only stream-copy cuts may miss requested edges; verify actual encoded boundaries or use an appropriate frame-accurate conform.

Verify the actual result [bl: invariant]

Bind review to the exact output hash or immutable revision and decision list. Listen to every changed speech join at normal speed with surrounding context, checking consonants, complete words, breaths, repeats, pops and accidental gaps. Watch sync and visible continuity at joins and across separately recorded audio. Then watch/listen to the complete rough story, including range interiors and opening/ending, for abandoned takes, repeated ideas, missing connectors and misleading order. Do not use coverage or effects to hide unresolved speech.

When a fresh word-timed transcript of the exact rough export is available, run the same trim audit on it in output time. Reconcile each remaining repeat with the intentional keeps, including findings inside retained shots. Listen for paraphrased duplicates the matcher misses. If a cut is repaired, verify the new export. An empty findings list alone cannot clear the dialogue.

Probe and decode the delivered file; reconcile streams, duration, cadence, source/output mapping, and intended cuts. Check for unexpected gaps, overlaps, black/frozen frames and audio drift. Decode success or a fresh transcript alone cannot prove naturalness or story continuity. Report what was actually heard, watched and mechanically checked. If listening, viewing or full-story coverage is unavailable, label that review incomplete, with affected decisions identified. A plan-only deliverable must not imply rendered QA. Changed bytes invalidate prior output-specific review; retime dependent captions after speech changes.

Once story timing is stable, hand off its exact identity and output clock to cutaways, music, sound, captions, color and export QA as needed. Later overlays preserve these timings; any retime requires a new version and affected downstream revalidation.

Return the editable decisions, selected/alternate-take rationale, rough export if produced, verification coverage and unresolved items. Separate mechanical readiness from the creator's story/taste approval. Do not call the cut clean while known accidental repetitions or unsafe word boundaries remain.

References

These inform the workflow; none is a runtime dependency or endorsement.

See package inspiration and license notes for scope and provenance.

Individual skills in this repo

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

Related Skills