hook-splitter

hook-splitter — one long recording → N tight standalone hooks

Cuts a composite into one MP4 per hook, removes every gap and flub, and delivers the whole set on a here.now gallery canvas where each video has its own comment box.

Built and proven on Rough Cut - Composite.mov (16:26, 21 hooks): 946s of hook spans became 627s across 21 videos, then three feedback rounds off the canvas.

The two rules that matter most

1. Never trim copy for pacing. Remove dead air, stumbles, aborted takes and repeats. Do NOT shorten what the speaker actually said to hit a runtime. If a hook runs long because the setup is long, ship it long and say so — the trim is the user's call. (Approved copy on a paid deliverable is never trimmed for pacing.)

2. Transcribe the RENDER, not just the source. This is the single highest-value step in the pipeline. Whisper smooths a repeated take into one clean line on a full-file pass, so real doubles are invisible in every source-side transcript — but they show up immediately in a transcript of the cut. Four of the flubs the reviewer caught on the proving run were of exactly this kind. scripts/qa.py exists to catch them before the reviewer does.

Pipeline

Everything is driven by one hooks.json (see assets/hooks.example.json). Scripts take its path as $1 and default to ./hooks.json.

#	Script	What it does
1	`probe.py`	reads the source, extracts one wav per audio stream, computes 10 ms dB envelopes, and tells you which stream is the mic and which is the system capture
2	`video_regions.py`	the "a video is playing" mask (see below)
3	`detect_loops.py`	where each played video restarts (players loop)
4	`build_mix.py`	`work/mixed.mov`: video stream-copied, audio tracks summed, loop tails muted
5	`plan.py`	`cuts.json` — speech map, gap collapse, your hook table + kills
6	`render.py`	`final/hook-NN.mp4`, one per hook
7	`qa.py`	re-transcribes every render and scans for repeats (never skip)
8	`build_canvas.py`	the gallery canvas → publish with `here-now`
9	`map_notes.py`	round 2+: pull the reviewer's notes back and map them to source time

The spine: a clean "video is playing" mask

A screen recording made with a mic plus system-audio capture has two audio streams, and the system one is digital silence whenever nothing is playing. That gives an exact, free mask, and the entire edit hangs off it:

inside a video region → never cut. That is the only silence the brief allows.
outside one → the speaker is talking; every gap collapses to a fixed residual.

probe.py identifies the streams by their silence fraction (the system track is

50% digital silence; the mic never is). Verify alignment before trusting it — transcribe the same 45s window from both near the start and near the end. It has been exact so far, but a drifting mask would corrupt all 21 outputs silently.

Only one audio stream? Then there is no free mask. Mark the video regions by hand in hooks.json ("video_regions": [[13.2,16.45], …]), reading them off scripts/probe.py --map output plus the transcript. Everything downstream is unchanged. Do not try to infer them from the mic track alone — the ad audio bleeds in ~22 dB down and sits right on the speech threshold.

Delivery

build_canvas.py + the here-now publish script. It is a 21-up gallery, not the single-video canvas: the comments collection carries a hook field so each video gets its own box, and a note stamps that player's current frame.

Reply LEADS with the canvas URL, every round (see video-review-canvas).

Bump version every round so answered notes stop cluttering, and version-stamp the filename of every hook whose bytes changed (hook-04-v2.mp4) so no one gets a cached old cut. Delete the superseded files; verify the old name 404s.

About this skill

The two rules that matter most

Pipeline

The spine: a clean "video is playing" mask

Delivery

Read next