Video Hook Forensics

Diagnose why the first 6 seconds of your short-form video loses viewers — then get a hook you can cut today.

What it does

Measures the opening of a video second by second with FFmpeg, names the specific reasons attention leaks, and returns:

1. Second-by-second evidence timeline — every second of the 0–6s window, what's on screen, measured or seen
2. Up to 3 named findings — each with a timestamp and the frame or measurement that proves it
3. 3 structurally different replacement hooks — each tied to a finding, each built from footage you actually have
4. A timecoded edit table — source timecode → new timecode → operation, ready for an editor to execute

Plus a machine-readable hook-forensics.json for pipeline use.

How it works

This measures cut positions, shot lengths, a frame-to-frame motion proxy, per-second audio loudness, and samples frames for the opening window (default: first 6 seconds, configurable). The skill then reads the sampled frames, applies a structured diagnostic framework (6 attention functions, 8 named failure modes, 7 rebuild structures), and writes the report.

Requirements

• FFmpeg and FFprobe on PATH (ffmpeg -version to check)
• No FFmpeg? The skill falls back to transcript + keyframes mode with technical metrics marked unknown — still useful, just less precise

Example finding

F8 Payoff leak — everything is delivered at 0.0s
Evidence: frame at 0.0s and frame at 5.0s are visually identical; price, rating, and CTA are all readable in the first frame.
Consequence: after second 0 the video has nothing left to give — seconds 1–6 are pure cost.

Example measured result (real before/after)

Built from the creator's own footage, measured with this same tool before and after — full methodology in the skill's before-after-protocol.md.

What this is not

• Not a full-video editor. It only analyzes the opening window (default 6s, configurable) — not pacing, thumbnails, or channel strategy.
• Not a virality predictor. It never claims a retention percentage, view count, or "this will go viral." It reports measured structure only — cuts, motion, loudness, timing — and separates what was measured from what is interpretation.
• Not a copy tool. Replacement hooks must not reuse the analyzed video's wording, shots, or branding — including when analyzing a competitor's video for research. Mechanisms (timing, structure, information order) are fair to study; expression is not.

Best for

• Diagnosing why a specific clip's opening isn't holding attention
• Rewriting a cold open before a re-cut
• Competitor opening teardowns (for research — not for copying)
• A/B hook planning with a structured, evidence-first process

FAQ

Does it need my video's real analytics?
No — it works from the video file alone. If you supply retention data, it will fold that in; without it, findings stay structural (cuts, timing, framing) rather than performance claims.

What if I don't have FFmpeg installed?
The skill tells you how to install it (Windows/macOS/Linux commands included), or continues in a reduced mode using a transcript and keyframes you supply.

Will it tell me my video will go viral?
No. It deliberately never makes that claim — see "What this is not" above.

Privacy / data handling

This Skill runs entirely on your own machine, inside your own Claude Code
(or compatible AI client) session. No video file, extracted frame, or
evidence JSON is uploaded to Capafy or any third-party service by this
Skill — all FFmpeg processing and analysis happen locally.