reel-style-clone

reel-style-clone — forensic style extraction from a reference reel

Take a reference reel apart with ffmpeg forensics — cut detection, contact sheets, audio spectrograms, whisper transcript — analyze the evidence frame-by-frame (with parallel subagents when available), and synthesize a style guide + per-line build directives that let a build skill reproduce the style on completely different footage.

The core belief: an editing style is not a vibe, it is a countable, nameable system — a cut density, a caption register grammar, a fixed graphic vocabulary, one signature trick. Extract the system and you can apply it to any script. Proven on a real style-clone delivery that went through client review rounds.

Read references/method.md before running the pipeline — it has the exact commands, the subagent prompts, the style-guide template, and the hard-won analysis heuristics. This file is the map; that file is the territory.

Inputs

1. The reference video, as a local file. The user provides it. If they only have a platform URL (Instagram/TikTok/YouTube), platform downloads usually require the user's own authenticated session — ask them to download it themselves (screen-record as a last resort) or to provide their logged-in browser session; do not scrape anonymously and expect it to work.
2. The user's own footage and/or script that the style will be applied to. (Can arrive later — phases 1–5 need only the reference.)

Pipeline

1. Probe & cut detection

ffprobe the reference (duration / resolution / fps). Detect hard cuts with select='gt(scene,0.25)',showinfo → a list of cut timestamps and a PNG of the first frame after every cut. Compute cuts/sec overall and per-section shot lengths (hook vs body vs CTA usually pace differently). Commands: references/method.md §2.

2. Frame evidence

Extract frames at 2fps scaled to 360w; tile them into 5×4 contact sheets (20 frames = 10 seconds of video per sheet). A 60s reel becomes ~6 images an agent can actually study. §3.

3. Audio evidence

Extract audio → whisper-transcribe (base model is fine — you need narration beats, not perfect words) → render a spectrogram (showspectrumpic) and waveform (showwavespic). From these, infer: music present or not, BPM from kick spacing, where SFX hits land, and the energy arc (where the mix lifts and drops). §4.

4. Parallel forensic analysis

When subagents are available, fan out — one agent per contact sheet (frame-by-frame: caption typography/position/case/colors, graphic overlay types, B-roll grammar, framing modes, grade), one agent on the cut frames (shot-alternation grammar — what each cut lands on, flash/transition tricks), one agent on the audio evidence. Solo fallback: same passes, sequential. Prompts: §5.

5. Synthesize the style guide

One synthesizer merges the reports into a single guide with exactly these sections — PACING / CAPTIONS / GRAPHIC LANGUAGE / B-ROLL GRAMMAR / COLOR & GRADE / SOUND DESIGN / BUILD DIRECTIVES. Sound design is written as an ElevenLabs music prompt so it is directly generatable. Build directives are a per-line plan applying the style to the USER'S script: cut points, visuals, caption treatment, SFX hits — matching the reference's cuts-per-second density. Template + rules: §6–7.

6. Hand off to the build

Give the style guide + build directives to branded-ad-edit (the full motion-graphics pipeline) — or reel-recut if the reference style is just banner + karaoke captions + tight pacing. The directives are written so the build skill can execute them without re-watching the reference.

The five analysis lessons (why this works)

These are the difference between "purple captions, fast cuts" and a guide that actually clones the style. Full detail in references/method.md §8.

1. Mode/register systems are the signature. Editors run conditional rules, not constant looks — e.g. sans-serif lowercase while the face is on screen, serif ALL-CAPS when a graphic replaces it. Catch the condition, not just the two looks, or the clone reads wrong instantly.
2. Count internal builds, not just cuts. Perceived pace ≈ 2× the hard-cut rate, because elements pop on within shots (caption words, list items, zoom punches). A clone that matches cuts/sec but not builds/sec feels sluggish next to the reference.
3. Find the once-per-video trick — and use it exactly once. Most refs have one spike moment (a light-leak flash, an inverted frame, a speed ramp). Used once it is a signature; repeated it is a gimmick and the clone feels cheaper than the original.
4. Name the recurring devices. e.g. screenshot-over-own-blur (a screenshot floated over a blurred blow-up of itself — fills 9:16 with any aspect evidence) and the blue reading-highlight (a translucent swipe that paints text as the VO reads it — eye control). Devices you can name, you can rebuild.
5. Say what each overlay is FOR, not just what it looks like. "White serif caps on black" is copying; "the AUTHORITY register — used when the edit wants you to take a claim as fact" is cloning. Every graphic in the guide gets a function, so build directives can deploy it on new lines the reference never spoke.

Outputs

Write into the project working directory (not into this skill):

• style-guide.md — the seven-section guide (§6 template)
• build-directives.md or a table inside the guide — per-line plan for the user's script
• evidence/ — cuts list + cut PNGs, contact sheets, spectrogram, waveform, transcript (keep these; review rounds re-open them)

Requirements

• ffmpeg / ffprobe on PATH
• A local whisper (whisper-cli + a ggml model — base is enough here), or npx hyperframes transcribe (node ≥ 20) which wraps it
• Subagents (Task tool) optional but strongly recommended for §5 fan-out
• No API keys needed for analysis. The downstream build via branded-ad-edit needs ELEVENLABS_API_KEY in a .env at the repo root — that's its requirement, documented there.