hook-cutter: 롱폼 → 점수·자막 Shorts

hook-cutter

Turns long-form video into captioned vertical Shorts/Reels (and optional 16:9 clips): find the hooks, make the cuts, add the captions, make the covers, prove it with QA. You supply the judgment: which moments, in what order, with what words on screen. The scripts do the mechanics deterministically and write inspectable JSON at every step. The skill ends at local files.

Per approved clip it makes: renders/<clip>__vertical__<style>.mp4 (1080x1920 H.264/AAC at -14 LUFS, word-timed captions with one accent word, a hook card readable on frame 0), an .srt, a 1080x1920 cover with grid and mobile previews, a QA report, and a named export with a manifest.

Hard rules

1. Never upload, post, schedule or publish. Gate E is the user approving local files.
2. Never attribute words to the wrong person. If attribution is not certain, drop the clip.
3. Never put words in anyone's mouth. No guessed words, no meaning-changing omits, no generated or cloned faces or voices. A garbled phrase gets cut around, never "fixed".
4. Every number and name on a written surface (hook card, cover, title, post, chapter title) is said in the clip. One exception: the clip's speakers may be named in the title and the post.
5. P/moments.md is the only status store. A row is CUT only when its render exists.
6. No virality scores. Ledger scores sort the shortlist; never present them as predictions.
7. A clean exit code proves nothing visual. Open every image QA lists under EYES.
8. Fix the spec, never the output. Never hand-edit renders or QA files; never loosen a gate.
9. Ask about rights first: does the user have rights to this footage, and whose brand does it carry? Never apply another organisation's branding.
10. The project folder P is durable (never /tmp). Everything for a source lives inside it.
11. Images cost context; old ones drop out. Open at most two at a time, write each up in a notes file at once, never re-open one written up (qa.md, "Image budget").

Setup

hc below means python3 <skill-dir>/scripts/hc, where <skill-dir> is the folder holding this file (usually ~/.claude/skills/hook-cutter).

hc doctor                          # READY table: ffmpeg, Python, venv, models, fonts
hc doctor --install                # once: venv + deps (uv builds 3.12; --python PATH for 3.13)
hc doctor --install --with url     # + yt-dlp for URL sources (YouTube also needs deno or node)
hc doctor --install --with mlx     # + faster ASR on Apple Silicon
hc doctor --model large-v3-turbo   # pre-download the ASR model (about 1.6 GB; --backend picks whose)

If any command prints hook-cutter is not set up, run the install line. Never pip install into the venv yourself.

A new user: offer hc demo first: one finished Short from a made-up recording in about a minute, through the same commands below (needs say or espeak-ng; WARNs such as "no face" are expected on its drawn speaker).

Workflow

Start with one message to the user covering: the source, the project folder, who the clips serve, rights and branding, anything already posted from this source, and (only if they care) count, formats and style. Then:

1. Ingest

hc ingest P SOURCE                 # local file (linked; --copy to copy) or URL (--section 12:00-48:30)

Read the warnings in P/source.json (low resolution, HDR, VFR, missing frames, audio holes, a channel far quieter than the other: then mixed to mono) and open P/sheets/survey.jpg: how many people, black bars or gallery tiles, layout changes. A letterboxed or gallery source gets video.active_area. Re-ingest a VFR source or one with missing frames with --cfr (the transcript stays valid). A source replaced after ingest stops every step until it is ingested again (spec.md, "Source identity"). Tell the user what caps quality before promising sharp verticals.

2. Transcribe

Write P/glossary.txt first: names, companies, jargon, one per line (from the user, the video title and description, the survey). Then:

hc transcribe P                    # --speakers-vtt PATH if a speaker-labelled VTT exists

Skim P/transcript.md for mangled proper nouns. The same mistake everywhere: add it to the glossary and run hc transcribe P --force. A single occurrence: fix it in clips.json later.

Transcribe drops words with no audible sound, re-times word edges from the audio and re-listens to voiced stretches that got no words. Flags in transcript.md: unworded_audio and dropout (<<... s of sound with no words>> lines), false_start (two attempts at a phrase, a visible restart such as "I was, I was" included), timing_uncertain (a start under a laugh or another voice), speaker_uncertain, no_audio. Check them with hc words before keeping that audio (hooks.md section 1). You cannot listen: qa.md ("Checking sound that has no words") shows how to transcribe one stretch alone; clip key insert captions speech the transcript missed. hc transcribe P --retime re-times without ASR (indices stay valid). HOOKCUTTER_DEVICE=cpu forces faster-whisper onto the CPU.

3. Mine the moments

Read references/hooks.md first, every time. Recall pass in 10-minute windows with 1-minute overlap into P/moments.md, then hard gates, scores, and the set pass. Tools: hc find P "phrase", hc words P A B, hc peek P START END, and P/signals.json for URL sources (heatmap and chapters: they surface candidates, they never rank them).

Gate A (user): approve the shortlist. Show the shortlist table (hooks.md section 9) and stop. Approved rows become PLANNED once they have a clips.json entry.

4. Look before you build

For every approved row:

hc peek P 24:00 24:40              # contact sheet of the span
hc shots P --range 24:00 24:40     # shot boundaries, faces, proposed framing

Gate B (agent, hard): attribution and on-camera. You are certain who says the hook line, and that person is on camera when it lands (or a wide shot honestly shows the group). Not certain: mark the row PASS attribution and tell the user. No exceptions.

5. Write the edit spec

Write P/clips.json: references/editing.md for the craft, references/spec.md for every key, examples/clips.example.json for a full example. Every piece carries expect anchors. Write hook_card and cover.headline per references/packaging.md.

Multi-face shots (a gallery call, a two-shot): hc shots prints MULTI-FACE and proposes letterbox; never paste a single-face mark there. Set "speakers": {"<label>": {"x": 0.27}} once in clips.json (x from the shots sheet, labels as in transcript.md): framing follows the speaker turns, and QA G16 fails a render that frames the wrong face. Then:

hc shots P c01-some-id             # shots for the clip's real range: framing marks come from here
hc plan P                          # piece table + lints; exits 1 on any FAIL lint

Framing needs face data. hc shots P <clip> writes P/shots/<clip>.json and draws P/sheets/shots_<clip>.jpg: LOOK at the sheet before you write a framing mark (a mark is your judgment of who is on screen). hc render and hc all analyse the shots themselves when they are missing, so a mark never falls back to a blind crop (the mechanics, not the looking).

Fix every FAIL. Read every WARN and fix it or note in one line why it stands. plan is cheap: run it after every edit.

6. Render

hc render P                        # all clips; --styles clean,pop  --formats vertical,landscape

When render prints analysed shots for <clip> (framing needs face data), nobody has looked at its shots yet: open P/sheets/shots_<clip>.jpg and check the marks before you trust the render.

Gate C (agent): read the sheets. For every render open P/sheets/<clip>__<fmt>__<style>.contact.jpg and P/sheets/<clip>__<fmt>.seams.jpg: the right person framed in every cell, captions clear of the mouth, the hook card readable in cell 1 (frame 0, the default thumbnail), no black band, black or flash frame or framing jump at a seam. Fix in clips.json (framing marks, speakers, caption_y, piece edges) and render again. Parallel hc all P <clip> runs on different clips are safe (shared files are written under locks).

7. Covers

hc cover P c01-some-id             # candidates sheet + cover + .grid.jpg + .mobile.jpg
hc cover P c01-some-id --at 12.4   # force a frame (output time: 12.4 or 0:12.4)
hc cover P --grid                  # only rebuild sheets/covers-grid.jpg from the covers on disk

Open P/sheets/<clip>.cover-candidates.jpg and the cover's .grid.jpg. Pick an expressive frame of the person who owns the claim; record it as cover.at_out in clips.json.

8. QA

hc qa P --require-coverage

Gate D (agent): zero FAIL, every EYES image opened. Read P/qa/summary.md and each P/qa/<clip>.md. Zero FAIL. Every WARN fixed or explained in one line. Open every image under EYES and list the ones you opened. SKIPPED is not PASS: a clip whose gates passed but some could not run is INCOMPLETE; a WARN clip with such gaps is held too (exit 2 under --require-coverage). Run this full QA (with ASR) before the final export. references/qa.md explains every gate.

9. Export and present

hc export P

Export holds back clips whose QA is FAIL, INCOMPLETE, stale, missing or has evidence gaps, and renders that no longer match their inputs (exit 1, listed); --force exports them anyway, only when the user says so. It also writes P/export/by-ear.md: every join, the opening, the ending, flagged words and sound WARNs, each with its time. Present: the P/export/ paths, by-ear.md (you cannot listen: ask the user to play every clip with it open), P/sheets/covers-grid.jpg, the QA summary (FAIL and WARN counts plus anything you explained), the three surfaces per clip (hook card, cover headline, title), draft post copy if asked, and any cherry-picks or risks.

Gate E (user): approval. Approval means the user has the files. You never upload or publish. Update moments.md: CUT rows get what was exported; anything deferred becomes BANKED with a reason for next time.

The iteration loop

User notes are one-liners ("open stronger", "the laugh got cut", "wrong face at 0:12"). Turn each into one spec edit (references/editing.md ends with a note-to-edit table), then:

hc all P c01-some-id               # plan, render, cover, qa, export; skips work whose inputs did not change
hc status P                        # one line per clip: plan / render / cover / qa state

• hc all passes --styles/--formats on, runs QA with --require-coverage (--no-asr leaves QA INCOMPLETE, so export holds everything: a quick look, never the final files), exports after a QA FAIL (holding failing clips), exits with the worst code and ends with one line per clip saying what ran. It re-makes only stale covers and keeps a current one made by hc cover or copied in. hc status judges freshness by content and names the reason: plan STALE(spec|shots|transcript|source|code), render STALE(plan|shots|spec|source|inputs) (inputs: preset, font, safe zone or render code), cover STALE(at_out|headline|shift|render).
• One note, one edit, one re-render. Do not touch clips the user did not mention.
• Style-only changes reuse the cut; edits to pieces re-cut only the segments that changed.
• Re-run gates C and D for every clip you re-rendered.
• Log notable outcomes (a reversed PASS, a reverted reorder) in the ledger for the next session.

Command reference

Command	Use
hc doctor [--install] [--with mlx,url,whispercpp] [--model NAME] [--backend auto|mlx|faster-whisper] [--python PATH] [--json]	check or set up the environment
hc ingest P SOURCE [--copy] [--section A-B] [--cfr] [--force]	bring a file or URL into a project
hc transcribe P [--backend auto|mlx|faster-whisper|whispercpp] [--model M] [--language L|auto] [--glossary TEXT] [--speakers-vtt PATH] [--force]	ASR to transcript.json and transcript.md
hc transcribe P --retime [--clip ID]	re-time existing words from the audio (no ASR)
hc transcribe P --file VIDEO --clip ID	transcript for a caption-only clip
hc find P "phrase" [--near MM:SS] [--max N] [--clip ID]	phrase to word indices and times
hc words P A B [--clip ID]	words A..B with times, gaps, filler flags, speakers
hc peek P START END [--n 12] [--file PATH]	contact sheet of a source range or a rendered file
hc shots P CLIP... / hc shots P --range START END	shots, faces, framing proposal and sheet
hc plan P [CLIP...] [--quiet]	validate and plan; prints table and lints
hc render P [CLIP...] [--styles a,b] [--formats vertical,landscape] [--force]	render clips and sheets
hc cover P CLIP... [--styles a,b] [--at SECONDS]	cover, candidates sheet, previews (--at takes any time form below)
hc cover P --grid	rebuild sheets/covers-grid.jpg from the covers on disk (seconds; after an interrupted cover step)
hc qa P [CLIP...] [--styles ...] [--formats ...] [--no-asr] [--require-coverage]	QA gates and reports
hc export P [CLIP...] [--styles ...] [--formats ...] [--force]	named files, export/manifest.json and export/by-ear.md; holds unchecked clips (after adding, dropping or reordering clips, export all clips)
hc all P [CLIP...] [--styles ...] [--formats ...] [--no-asr] [--force]	the whole chain, skipping up-to-date work (--force redoes every step; held exports stay held)
hc status P	where every clip stands
hc test [--slow] [PATTERN]	the skill's own test suite
hc demo [DIR] [--force] [--open]	one finished Short from a made-up recording, to see the output (default ~/hook-cutter-demo)

Times accept 681.5, 11:21, 11:21.40 or 1:02:03.4. Default clips = all clips in clips.json; default styles and formats = the clips.json defaults.

Exit codes: 0 ok, 1 spec or user error, a QA FAIL, a held export or a usage mistake (the message says what to fix), 2 evidence gap, 3 tool or environment error such as a read-only folder or a full disk (run hc doctor). HC_VERBOSE=1 prints ffmpeg command lines.

References

File	Read it when
references/hooks.md	before mining: recall pass, ledger format, taxonomy, gates, scores, set pass, shortlist, worked examples
references/editing.md	writing pieces: cold opens, preamble, repetition, endings, length bands, notes-to-edits
references/captions.md	choosing a style, accents, legibility, safe zones, hook card mechanics, glyphs
references/packaging.md	hook card vs title vs cover, parse-cold covers, frame curation, post copy
references/fidelity.md	anything touching attribution, numbers, risky claims, rights
references/qa.md	reading plan lints and QA gates, EYES discipline
references/spec.md	every clips.json and preset key with defaults and examples