Media-Editor-MCP: Skill for Claude Code

Create reel (each clip once) — natural-length beat edit

A variant of [[pp-create-reel]] for when you DON'T want to repeat footage to fill 59s. Each clip plays once; the reel ends when the clips run out (e.g. 5 clips ≈ 10-25s depending on spans). Cuts still land on downbeats. Verified tooling is identical — only the slot-allocation logic differs.

Inputs: music, footage folder, plus two knobs:

• order: chronological (clips in folder/given order — story order) or smart (open on the highest-scoring clip as a hook, rest by motion). Default chronological for this mode (it's usually a "show my clips in order" use case).
• beatsPerClip: how many downbeat-intervals each clip spans — 1, 2, 3, or auto (default). auto gives longer/richer source clips more beats (see Step 3). Spanning ≥2 beats gives a clip more screen time and suits cinematic / slow footage.

This mode does NOT force 59s, does NOT repeat clips, and has NO loop ending (each clip is unique — looping would require repeating the first clip). Everything else (vocal-skip, smart slices, dissolves, grade, punch-ins, title) works the same.

Pre-flight & footage

Same as [[pp-create-reel]] Step 2: premiere_health + analysis_health green; probe footage and reject any clip with an audio stream; one orientation. NO ~12-clips rule here — use exactly the clips the user has.

Step 1 — Analyze

1. Vocal-skip (default on): transcribe → first vocal; music start S = first downbeat ≥ first_voice − 0.15 (instrumental → S=0).
2. detect_beats → downbeats (source time). Index the downbeats at/after S as D = [D0, D1, …] (D0 = first downbeat ≥ S). bar = median(diff(D)).

Step 2 — Order the clips

• chronological: keep the folder's natural sort / the order the user listed.
• smart: run find_best_moments on each, put the single highest-scoring clip first (hook), then the rest by descending top-window motion.

Step 3 — Assign a beat-span to each clip

Per clip, decide span ∈ {1,2,3} = how many consecutive downbeat-intervals it occupies:

• fixed (beatsPerClip=1|2|3): every clip gets that span.
• manual (beatsPerClip=[3,1,2,…]): one span per clip, in order.
• auto (default): a two-layer model — the SONG sets the cutting pace, then per-clip ACTIVITY distributes spans around it. This makes energetic tracks cut fast and cinematic tracks linger, while still giving the most dynamic clips the longer holds.
  1. Song layer — classify_song(music) → recommendedSpanCenter (energetic ≈ 1.3, balanced ≈ 1.8, cinematic ≈ 2.5) from tempo + onset density + percussive fraction. This is the average beats-per-clip.
  2. Clip layer — find_best_moments returns absolute, cross-clip-comparable activityMotion / activitySharpness (whole-clip means, before per-clip normalization). Min-max them across the clip set; activity = 0.7·motionN + 0.3·sharpN, giving each clip an activityRank ∈ [0,1].
  3. Combine: span = round(spanCenter + (activityRank − 0.5) × 1.5), clamped to [1,3]. (So on a cinematic song every clip leans 2-3 but the liveliest still top out at 3; on an energetic song most clips are 1 and only the most dynamic reach 2.)
  4. Cap by what the clip can fill: shrink span while clipDuration < span×bar + 0.3. A short clip can't span 3 even if lively; a long static clip stays low.
  • This replaces the old raw-duration rule (which wrongly gave long STATIC clips 3 beats). beatsPerClip (fixed/manual) still overrides everything.
• A clip's slot duration = D[k+span] − D[k] (sum of its bars). Get its slice from find_best_moments(clip, window_seconds = slotDur + 0.1, count=1) → top window start. (In chronological mode this still picks the BEST window within the clip — order refers to clip sequence, not within-clip.)