jsath/claude-skills

Edit style: dynamic vertical reels from a raw take

You turn one raw talking-head clip into a finished 1080x1920 reel that looks hand-edited: the frame never sits still, graphics are real UI or official logos (never stock, never invented), and captions land on the beat. Nothing is dragged on a timeline. Every decision is a written shot plan rendered by code, then checked by gates that measure the final file.

Part of a pipeline: use-local-whisper (transcription) -> edit-style (short-form edit) or longform-edit (YouTube long-form) -> reel-captions (post text) -> thumbnail-generator (covers).

Requirements

Claude Code on a computer (not Claude.ai: it needs a shell and ffmpeg).

need	for	install
ffmpeg + ffprobe (with ebur128)	every cut, composite, gate	brew install ffmpeg / apt install ffmpeg
Python 3.9+ and Pillow	captions, cue alignment	pip install pillow
a local Whisper	word timestamps	see the use-local-whisper skill
OpenCV (optional)	framing gate, crop solver	pip install opencv-python-headless + YuNet model (see tools/face_box.py)
Node + Playwright (optional)	headless screen capture	npm i playwright && npx playwright install chromium
Remotion (optional)	animated graphic panels	npx create-video@latest; a full panel studio is in github.com/johannships/reels-engine
swiftc, macOS only (optional)	iPhone HDR to SDR	xcode-select --install

Install check: sh tools/check_setup.sh (exit 0 = ready).

Inputs

• raw.mov (or mp4): one take, phone or camera, any orientation.
• script.txt: what was said, spelled the way it must appear on screen. If there is no script yet, write one first with references/script-writing.md.
• Optional: a topic URL (repo, product page) to capture, and sources for every number.

Pipeline (one folder per reel: every file for this reel lives inside it)

transcribe.py below is use-local-whisper/scripts/transcribe.py. Parallel edits never share a work folder: one reel, one folder.

1. Check the source. ffprobe -v error -select_streams v:0 -show_entries stream=color_range,color_transfer,r_frame_rate -of csv=p=0 raw.mov arib-std-b67 = iPhone HDR, tone map first; pc,bt709 = full-range camera file. Exact handling: references/source-footage.md.
2. Cut retakes, then pauses. Transcribe raw (transcribe.py raw.mov), keep the LAST take of any repeated line, never cut mid-word. Then python3 tools/tighten.py cut.mp4 tight.mp4 (-50 dB, never -38 dB).
3. Transcribe the tightened take: transcribe.py tight.mp4 gives tight.words.json.
4. Write the shot plan (shots.json): one row per shot, {layout, anchor_phrase, content}. anchor_phrase is script text matched against words.json so the shot starts when those words are said. Never hand-type timestamps. Layout rules: references/edit-craft.md.
5. Capture and fetch assets. Screens: node tools/capture_clock.js <url> shots/s3 6 (headless, frame-exact). Logos: the brand's official press page or site SVG only. Open every asset and look at it before use.
6. Frame the face per take: python3 tools/face_box.py tight.mp4 --model yunet.onnx --solve full (and --solve pip). Never reuse crops from another take.
7. Render panels (Remotion or stills) to each shot's exact duration, composite per shot with ffmpeg, concat, lay the continuous voice on top. Run renders under nice -n 15. Look-dev on stills, then one full render.
8. Mix and master (references/edit-craft.md section 9): high-pass 75 Hz, light denoise, synthesised SFX 12 to 20 dB under the voice, -14 LUFS, true peak at or under -1.5 dBTP.
9. Captions, timed on the FINAL audio: transcribe.py final_nocaps.mp4, then python3 tools/make_cues.py script.txt final_nocaps.words.json --fix fix.json, then python3 tools/burn_captions.py final_nocaps.mp4 cues.json final.mp4 --caps. Text comes from the script, timing from Whisper. --y 0.38 on PIP shots.
10. Gates (below). Fix and re-run until all pass.

Layouts

layout	what	when
FULL	face fills the frame (head-top 13%, chin 85%)	hook, personal beats, close
GFX	full-screen real UI, stat or claim card, no face	each key beat, the CTA
PIP	graphic on top, face in a rounded card at the bottom (y 1086, 1080x834, radius 48)	naming a tool while still talking

Switch every 2 to 5 s. Open on FULL or PIP, break to GFX on each key beat, land the CTA on GFX, close on FULL. If the topic is a named brand, show its official logo in the first 2 s.

QA gates (all must pass; write results to QA.md)

#	gate	tool	pass
1	framing	face_box.py final.mp4	head-top > 3%, chin < 94% on every sample
2	caption sync per chapter, on the final render	re-transcribe final, caption_sync_gate.py captions.json final.words.json chapters.json	median <= 0.12 s, p95 <= 0.25 s
3	A/V length	final_gates.sh final.mp4	within one frame
4	loudness	final_gates.sh	-14 +/- 1 LUFS, true peak <= -1.5 dBTP
5	colour/encode	final_gates.sh	yuv420p, tv range, BT.709 tags, >= 8 Mbps
6	skin luma vs source	skin_luma_check.sh	within 2 levels
7	first word survives	re-transcribe first 2 s	first script word present
8	contact sheet	final_gates.sh writes contact.png	you OPENED it: no pills or stickers, no caption over the mouth, logos correct, nothing private, every label inside the 4:5 crop (y 285 to 1635)
9	claims	manual	every on-screen number has a named source checked live today; false spoken lines are cut and listed

On a caption-sync fail: retime that chapter's cues to the final words, re-burn, re-transcribe, re-gate. Loop until every chapter passes.

Output

reel-folder/
  final.mp4        1080x1920, H.264 BT.709, AAC 48 kHz, -14 LUFS
  captions.srt     from the re-transcription of final.mp4, names spelled right
  contact.png      12-frame sheet you looked at
  QA.md            one line per gate, plus CLAIMS and any cuts of private detail
  LOG.md           written as you go

Example QA.md:

framing      12/12 PASS (head-top min 6.1%, chin max 88.4%)
caption sync hook median 0.04 p95 0.11 PASS | body 0.05/0.14 PASS | cta 0.03/0.09 PASS
A/V          diff 0.012s PASS     loudness -14.1 LUFS, -1.8 dBTP PASS
colour       yuv420p tv bt709 13.9 Mbps PASS   skin luma diff 1.2 PASS
contact      looked: no pills, captions clear of mouth, logos official, 4:5 safe
CLAIMS       "40,000 stars": GitHub API 14:05 today = 41,212 (shown as 41K)

Failure handling

• No Whisper: stop and install one (use-local-whisper). Never guess timings.
• No OpenCV: skip gate 1's tool but eyeball every contact-sheet frame for chin/forehead cuts, and say so in QA.md.
• No Remotion: build GFX shots as still PNGs (Pillow or a headless Playwright screenshot of an HTML card) with a slow push-in (zoompan).
• make_cues.py exits 2 (script and audio differ): fix the script to match what was said. Never burn captions that disagree with the voice.
• A gate fails three loops in a row: stop, report the numbers, ask.

Rules that keep it from looking like AI

• No floating pills, badges, corner tags, stickers or REC chips. Ever.
• No code-generated motion-graphics sizzle. Real footage and real product screens.
• One accent colour. One caption style per shot type.
• Show, don't narrate: when the voice names a thing, that thing is on screen.
• Nothing private on screen: finances, client names, account ids, internal files.
• No em or en dashes in scripts or captions.

Depth: references/script-writing.md (hooks, register, numbers), references/source-footage.md (HDR, full-range cameras, clone takes, silence), references/edit-craft.md (captions, screens, framing, gestures, audio, pitfalls).