---

name: music-edit description: >- Design, assemble, and refine music videos, fan edits, hype reels, and rhythmic supercuts from footage, a song, and optional style references. Use for creative edit direction, reference breakdown, shot selection, timing, transitions, and render-review iteration across FFmpeg, Resolve/Fusion, or custom compositing. Includes an optional indexed-corpus beat-edit pipeline; neither an index nor that renderer is required for editorial work. For talking-head cleanup use cut-talking-takes; for operating Resolve use davinci-resolve.

music-edit

Turn the user's footage, music, and references into an authored edit and a reviewable render. Choose tools after understanding the visual and musical intent. Reuse analysis and rendering helpers where they fit; their presets are implementation options, not the creative brief.

Read the reference as a sequence

When references exist, identify the passage that best expresses what the user likes. Review at normal speed for rhythm, then inspect intricate passages in slow motion and frame by frame. Look for subject and camera movement, incoming/outgoing poses, retiming, brief overlays, mask reveals, layer ordering, typography changes, color, motion blur, and musical accents. Record useful timestamps and distinguish observed behavior from a guessed technique. A contact sheet helps find passages but can miss the few frames that make a transition work. Use timestamped original-frame extraction when playback controls are insufficient.

Use available playback and audio/video inspection tools. Be precise about the evidence: viewing extracted frames is not watching the complete sequence, and reading a beatgrid is not hearing the music. When a modality is unavailable, continue with timestamped frames and signal analysis, state that review limit, and deliver a playable preview for the user's audiovisual judgment.

Keep the brief in the workspace: source/song window, intended feeling or arc, reference passages, and delivery aspect. Infer reasonable defaults from the material and user context; do not make a questionnaire a prerequisite. For a reusable brief with subject and media placeholders, adapt the music-edit prompt. Its high-energy, FX-heavy direction is a template choice, not a requirement for every edit.

Make editorial choices explicitly

Choose source ranges for the gesture, expression, composition, or continuity that makes the sequence work. For a large library, use footage-index to find candidates and then inspect the actual ranges, including transition handles. Ranks and representative frames narrow the search; they do not choose the cut. Neither do measurements: a shot that scores well on brightness, saturation or motion can still be the wrong subject entirely, and a single frame cannot show that the character walks out of the shot. Confirm each pick as a frame sequence across the duration the slot will actually use. For a handful of supplied clips, inspect them directly without indexing a whole corpus.

Use the song's phrasing, lyrics, attacks, pauses, and energy as cues. Beat analysis supplies timing evidence; the editor decides where to cut, hold, anticipate, or delay. A shot can cross several beats. An intentional offset belongs in the authored timeline; accidental render drift remains a defect.

Author effects around the chosen images and intended movement. A transition may depend on matching poses, motion direction, a foreground mask, or a graphic shape. Select or build that mechanism instead of choosing the nearest preset name. Effect count, randomness, and a louder profile do not measure quality. Keep source footage, generated graphics, masks, and their provenance identifiable.

A match cut is one such mechanism: the incoming subject occupies the screen region, size, and direction of travel the outgoing subject just had, so a cut between two films reads as one continuous move, and one or two masking frames at the join (isotropic blur, a black/white threshold frame, a whip) hide the residue. The alignment is the work; the masking frames only finish it. When the corpus has footage-silhouette samples, footage-matchcut ranks in-points for a chosen outgoing frame and edit-plan-action --matchcut chains a whole passage that way, emitting trans-blur-* and trans-thresh-out on the cuts that qualify. Judge the paired frames and the rendered join; the score only orders candidates, and a flash or blur frame in a reference is not the frame to match against.

Preserve the intended source motion

Judge subject movement separately from animated zooms, text, and masks. Inspect the source's own cadence and any baked-in slow motion before adding retiming. A 60 fps export does not guarantee fluid action: 24 fps footage sampled at half speed supplies about 12 new source poses per second. Choose retiming per shot; compare frame repetition, interpolation artifacts, and an alternative speed or source range in a short proof. Holds and stepped motion remain valid authored choices; they should serve the phrase rather than appear accidentally.

During an overlap, advance each moving source on its own time map with enough source handles. Sampling the incoming shot at the fixed cut boundary, or clamping every pre-cut sample to its first frame, freezes it until the reveal ends. Check motion before, during, and after the join, including source-time continuity at the handoff. Make intended freezes explicit in the authoring state. For a custom compositor, check its sampled source times as well as the render: camera movement can make every output frame different while the subject remains the same pose.

Choose the implementation that fits

Paths can be combined: index discovery, song analysis, selected-range processing, and a Resolve/Fusion finish need not share one planner or schema. Preserve source identity and timing at each handoff. Probe the installed capabilities before treating an old API limitation as a creative limit.

The bundled FFmpeg renderer has kind-based motion, section grades, and a finite FX vocabulary. Use it when those semantics fit. An authored Resolve timeline or composition does not need to be translated into its slots just to pass its lint. New reusable implementation belongs with the owning skill; workspace files own this edit's timing, assets, masks, compositions, and intent.

Review the result against the intent

For an unfamiliar style or custom transition, first render a representative short passage containing the difficult move and enough lead-in and recovery to judge it. Compare it with the selected reference: subject readability, motion continuity, transition shape, text legibility, and musical timing. Revise the largest visible mismatch before extending that treatment. The passage length comes from the phrase or move being tested.

Review the complete final render uninterrupted at a confirmed 1× speed with audio when available. Judge musical feel and sustained subject movement, then inspect intricate effects in slow motion and frame by frame at delivery aspect. Compare the fine detail with the reference and finish with another 1× pass; slow review can hide a stop-start rhythm. Auto-framing can follow bright background objects; check the subject through motion, not only at the midpoint. If playback stutters, check the player's actual rate and dropped-frame evidence where exposed, then compare a lighter preview and the export's timestamps with its source-time mapping. Regular timestamps alone cannot distinguish repeated poses from fluid action. Reset any review player you control to 1× for handoff. Keep technical verification separate from creative judgment: source bounds, frame count, mux sync, HDR conversion, and active effects can all pass while an edit still misses the reference. State the reviewed version, playback rate, whether continuous motion and audio were actually perceived, and any remaining review limit. Starting a player or taking intermittent screenshots does not establish a complete audiovisual review; use the fallback described above.

Preserve prior versions for comparison. Deliver a playable preview and the requested full-resolution output, plus the editable project or composition when that is part of the chosen workflow. Retain enough authoring state to reproduce and revise the result; planner argv alone does not capture manual changes.