Reel edit (editorial style, cheap path)
A repeatable face-cam reel pipeline. Do not rebuild anything the script or template already does.
R=~/.claude/skills/reel-edit/scripts/reel.py
P=./reels/reel-NN # one project per reel
Token rules (why this skill exists)
- Do NOT load the
hyperframes/hyperframes-cli/media-useskills unless something here fails or the user wants a look that isn't covered below. The template already follows the render contract. - Never
catthe big reference files. Copy a pattern from~/.claude/skills/reel-edit/examples/<scene>.htmlonly when you need that exact scene (see list below), and read just that one file. - Look at frames with
reel.py snap(a 270 px-wide strip), never full 1080x1920 PNGs, never by rendering. - Render the full reel once at the end. For music, give the user mp3 previews (
beds), not renders. - Use the transcript only for timing. Its spelling is wrong for Hinglish (gunna, Chatjee); on-screen words come from the script.
Pipeline
python3 $R new $P # template: index.html, type/card/statement scenes, fonts; downloads logos + sfx
python3 $R prep $P "/path/IMG_1.MOV" [more clips] # HLG->SDR (swift), 1080x1920 30fps, cinematic grade; voice stays RAW
python3 ~/.claude/skills/reel-edit/scripts/recut.py $P --speed 1.05 --maxtail 0.7 # gentle cut: whole sentences (recommended)
python3 $R words $P # Whisper on the CUT voice, onset-snapped -> edit/words.json (prints phrases)
python3 ~/.claude/skills/reel-edit/scripts/logos.py $P gemini="Google Gemini|Google LLC" site=https://example.com # real logos
python3 $R cut $P is the older tight cut (removes micro pauses, 1.12x). It is faster but can clip soft word endings;
prefer recut.py unless the speaker wants a very punchy rhythm.
- Pass raw clips from wherever they are. Several clips = joined in the order given.
- Words. Whisper usually returns good word times (use them as they are). For mixed-language speech such as Hinglish it sometimes translates to English instead;
wordsthen prints a WARNING. In that case write what was actually said (e.g. Roman Hinglish, with punctuation at each phrase break; the English shows you what was said) toedit/script.txtand runwords $P --script $P/edit/script.txt. That aligns the script to the real pauses and onsets, accurate to about 0.1 s on most phrases but up to about 0.6 s on rapid lists ("caption, reply, DMs"). Snap those moments and nudgedata-tby hand. - Flubbed take / repeated line: read the
wordsoutput, thencut $P --drop 3.1-5.4with the times as printed (drops add up across runs;--resetclears them). Then rerunwords. - Retakes hide inside one clip. People repeat a line without stopping the recording, and Whisper often merges or skips the repeat. Slice the CUT
assets/voice.wavinto 1-2 s bursts and transcribe each one (-l en --prompt "<script>"keeps Hinglish in Roman script), then drop the weaker say of each repeated line. cutoptions:--speed(1.12 default; 1.0 for a slow talker),--th(speech dB),--gap(pauses shorter than this stay inside a phrase, 0.14),--pre/--post(kept around each phrase, 0.05 / 0.07),--dip(micro-pause cut, dB above--th, default 0),--tail(hold after last word).- The auto threshold is
max(-30.5, noise floor + 4, speech level - 10), so room noise in the pauses is not mistaken for speech. Aftercut, look at the energy envelope: any stretch longer than ~0.15 s that sits well below speech level is a pause the viewer will feel. Tighten with--thrather than a higher--speed. - After
cut, every element withdata-fullgets the reel length, and the jump-cut punch toggles are written between// CUTS:BEGIN/END.
Build the edit (the creative part)
Work from edit/words.json + the script/beat sheet.
- type.html - one
.grpper phrase while the face is on screen. Small lowercase Inter line + one huge Instrument Serif italic keyword (170-240 px; shorter word = bigger). Each.wgetsdata-t= its word start. Group start = first word, group end = next group's start. Bottom of text at y 1290.g1is the cover frame: visible at frame 0 with no reveal. Last groupid="gLast"(no fade, loops). Extras already styled:.acc(coral word),.badge(white logo tile),.pair(logo × logo),.strike(coral strike-through). - Cutaways (graphics over the face, 1.5-6 s each, while a proof is being said). Host div in index.html with root-timeline start/duration; inside, times are LOCAL. Duplicate a scene file + host for each use and give each a unique
data-composition-id, ids and timeline key.card.html- paper#f3efe8, REAL screenshot in a white card, 3D tilt in, push-in, coral ellipse draws around the proof.statement.html- black#0e0e0e, real logo, 2-3 stacked serif words, coral tag line.- Patterns in
examples/:viral.htmlIG reel grid → ranking row + counter rolling to 157x ·hooks.htmlthree hook cards dealt + score bar to 100 ·skills.htmlgrid of skill tiles ·human.htmlrobotic text crossed out ·github.html= card ·trio.html= statement ·type.htmlincl. the typed "Add a comment… SKILL" box.
- Face motion in index.html:
drift(t, d, 1.0, 1.06)on every face stretch between cutaways;punch(t, 1.06, 1.16)on 1-2 key words; flash + shutter for "screenshot". - SFX (in
assets/sfx): whoosh 0.22 at each cutaway entry (−0.05 s), pop 0.35-0.4 on badges/cards, tick 0.45 on counters, sfx_003 0.45 on clicks/cards, shutter 0.6 on flash. Same track index can't overlap - alternate 7-10.
Editing rules (learned from real redos)
These came from edits a creator sent back. Following them avoids the most common "redo this" notes.
Cut
- Never cut a sentence short. Keep soft endings ("-s", trailing vowels) and the pauses inside a sentence; only remove
dead air between sentences and the retakes. Speed 1.05 at most.
recut.pydoes this;--add/--cut/--droptake SOURCE seconds (one comma list per flag). - Before building, transcribe every kept span on its own and check its last word. If a take trails off, use another take.
- Fade every join (40 ms in / 80 ms out). Extend an end only while the voice is still going; stop at real silence
(
--maxtaildoes this) and never run into the next sound. - If something was playing in the room while recording (a reference video, music), keep the speech threshold above its level so fragments of it never survive the cut.
- Long Whisper runs on mixed-language speech can loop on one line. Transcribe 25 s chunks or each span separately.
Look
6. Keep the voice raw (gain + limiter only) unless the speaker asks for processing.
7. Real logos only: the product's official app icon (logos.py uses the App Store's artwork) or the product site's own
icon/vector logo. Never a placeholder letter tile, never the parent company's logo for a product, never an upscaled
48 px favicon. Look at assets/logos/app/_sheet.png before building.
8. When a name is spoken, its logo and name pop in on that word. For lists, time names from per-sentence transcripts
mapped onto the cut; a whole-reel alignment drifts by about a second.
9. Don't put small joiner words ("and", "or") between names in a big header: header = number + category, names under the logos.
10. Only web fonts render reliably: Inter and Instrument Serif ship in assets/fonts. System fonts and Apple Color Emoji
fail check; draw emoji with Twemoji SVGs.
11. Every claim on screen must be true on the day you post. Put limits on screen ("50 free credits/day"); if a price
can't be confirmed, don't call it free.
Optional layout: face-down window
For proof-heavy reels, keep the hook full-frame, then slide the face video down into a rounded window in the lower half and put
titles + proof cards above it. A cut-out of the speaker on top lets the hair or hands break over the window edge (optional: many creators prefer the
plain video in the window, which is faster and keeps their real background; then hide the cut-out layer).
scripts/lib_v2.py has the shared pieces for this layout: window move, header type groups, word reveals, SFX tracks.
- Cut-out:
npx --yes [email protected] remove-background assets/face.mp4 -o assets/face_cut.webm --quality best(runs locally, roughly 6 min for 14 s on Apple silicon). Give the cut-out clipdata-start="0" data-media-start="0"so it decodes in sync. - Layers, bottom to top: dark stage background < face window (base video inside a
clip-path: inset(...) round ...wrapper) < proof cards < cut-out < captions. Two face groups (window, cut-out) get the same slide/scale tweens; keep the cut-out hidden while the face is full-frame and fade its top ~80 source px with amask-imagegradient so a head cropped by the camera does not show a hard edge. - Close-up footage needs the window small (scale ~0.6, centred, top of the window around y 870) so the whole head and shoulders fit. Don't punch-zoom inside the window; jump-cut punches only while full-frame.
- Captions become titles above the cards (bottom of text ~y 410) while the face is down, and go back to y 1290 when full-frame.
Real assets only
- Logos:
scripts/logos.pyfirst (official app icons / the product site's own logo). For simple monochrome marks you can also usenpx --yes [email protected] media-use resolve --type logo --entity "<Name>"(thesvg). If it returns a low-res .ico (ChatGPT did), use simple-icons:https://cdn.jsdelivr.net/npm/[email protected]/icons/<slug>.svg.newdownloads claude/instagram/openai/github from simple-icons. - Web pages (GitHub, docs, product sites): screenshot with the cached chrome-headless-shell at deviceScaleFactor 2.7 →
assets/shots/. Delete the placeholdershot.png. - Instagram thumbnails / view counts: read from the logged-in built-in browser. Blur other creators' handles.
- Never invent numbers on screen; every stat must match a real screenshot.
Check, look, render
cd $P && npx --yes [email protected] check # must pass; fix overlap/occlusion errors it lists
python3 $R snap $P 0.5,4.2,7.9,13.4 # strip of key moments -> edit/snap.jpg (Read it)
python3 $R render $P reel-NN # -> renders/reel-NN.mp4 (SDR, 30 fps, high)
Snap the cover frame, one moment per cutaway, and the end. Fix, re-snap only what changed.
Music & delivery
- Default delivery: voice + SFX only (music line commented out). Add a trending track in the Instagram app when posting.
- If they want a baked bed: upbeat lo-fi / chill tech. Use tracks you have the rights to (CC0 / CC BY).
python3 $R beds $P track1.mp3 track2.mp3 ...→ mp3 previews to pick from; thenpython3 $R music $P <track> [--start S]→assets/music.wav(ducked under voice), uncomment the<audio id="music">line, and put the CC BY credit in the caption.
Gotchas
- iPhone clips are HLG HDR:
preptone-maps them; if a render comes out HDR/washed, the face file lost its bt709 tags. - Voice:
prepreads the audio straight from the original clip and keeps it raw; the cut levels it to -14 LUFS with plain gain + a limiter. If a speaker wants a brighter sound, add EQ yourself and let them approve it by ear first. - A GSAP
fromToplaced later on an element that is visible earlier needsimmediateRender: false, or its from-state shows from frame 0. - Inter and Instrument Serif load from
assets/fonts(each composition declares its own@font-face); never rely on system fonts. - Every timed element needs
class="clip"; a scene's timeline must be registered aswindow.__timelines["<composition-id>"]. - Text safe zone: y 230-1440, keep the right 230 px clear of key text (IG buttons).