Edit Style — dynamic vertical reels from a raw take

You turn one raw talking-head clip (an AI clone render OR a real recording) into a finished 1080x1920 reel that looks hand-edited: the frame never sits still, graphics are real UI, real logos or generated (not stock), and captions land on the beat.

Nothing here is dragged on a timeline. Every decision is written as a shot plan and rendered by code.

Keep a working toolkit folder — a STYLE.md (your measured spec), the pipeline scripts, the Remotion panels, and past scripts with their verified facts. Start there rather than rebuilding. One-command engine for the house formats: the house-reel-edit skill (skills/house-reel-edit). Small gate and capture scripts: tools/ next to this file.

The stack (all free/local except the clone)

• Whisper — word-level timestamps. This is what makes captions and cuts land on the audio. whisper-cli + ggml-small.en.bin for the transcript and SRT text; mlx-whisper large-v3-turbo for caption cue TIMING (small.en drifts 0.1–0.4s on the same audio, measured 2 Oct 2026).
• Remotion (React → video) — every graphic panel rendered to an exact duration. Runs standalone by symlinking this repo's studio/node_modules.
• Playwright, headless only — real screen footage (see Screens).
• ffmpeg — crops, composites the layouts, burns captions, mixes voice + SFX. NOTE: a stripped ffmpeg build may have — no drawtext, subtitles, libass. Text is rendered to transparent PNGs with Pillow and composited with overlay.
• HeyGen — only if the take is an AI clone.
• Claude Code — orchestrates all of the above.

Part 1 — Writing the script

The edit can be perfect and the video still reads as AI if the words are wrong. This is the part that gets rejected most often.

Beat structure

Roughly 140–155 words, landing 32–40s at 3.8–4.0 words/sec.

The hook

The first three words decide it. Openers that work, in order:

1. Second person — "Your AI agents…" — the subject is something they own.
2. The product's own name when it sounds impossible — "Free Claude Code…".
3. A recognisable brand — "Cash App's parent…".

Weak first-three-words: anything that starts with a concept, a preamble, or the word "There". Never open with a rhetorical question.

Register — how to not sound like AI

Two scripts were rejected outright for this. The tells were not formality. They were performed style:

• Never the "Not X. Y." antithesis. "Not as a bot. As an actual member." This is a copywriting tic and it is the single most recognisable AI tell.
• No performed asides. "and yeah, that Block. Square, Cash App." Trying to sound offhand reads as more AI, not less.
• No technical-writer register. "the difference isn't cosmetic", "by a factor of forty", "is a signed event in one log".
• No tacked-on trailing clauses. "…, in the same place your team talks."
• Raw numbers, never ratios. "Their next biggest repo has 700" beats "by a factor of forty". Humans quote the number and let you do the maths.
• Concrete comparisons over jargon. "you give it access the way you'd give a new hire access" beats "scoped by identity, not permission flags".
• No em-dashes or en-dashes in any script or caption. Standing rule, verbatim: "Remove any em dashes or hyphens, people know it's AI." Use a period, a comma, or rewrite the line. This is the same tell as everything above and it is the one most likely to slip through.

What actually works: plain declaratives, contractions, the occasional fragment for emphasis, and one first-person judgement ("that matters because…"). Say the thing and stop. Real speech isn't stylised, it's direct.

Two accuracy rules that bite

• Write numbers as digits — 2,000, 700,000, 47,000. Whisper emits "700,000" as three tokens, so a script saying "seven hundred thousand" fails to align and captions silently drop words. Digits also read faster on screen.
• Never claim a repo "launched today" unless you checked created_at. Trending ≠ new, and the creation date is two clicks away. Attribute unverifiable vendor claims ("the repo puts it at…") rather than asserting.

Part 2 — The voice

Default is a raw take from the creator; the agent edits. Clone only as a fallback.

AI clone takes

• Pair clone N with voice N. Mismatched pairs sound wrong. Current production pairing is 11 + 11 (owner revert 31 Aug 2026; clone 13 rejected in production). Speed 1.5, ear-approved 12 Sep 2026.
• SPEED 1.45–1.5. Slower reads as AI. 1.2 was rejected as "way too slow". 1.5 is HeyGen's ceiling.
• One continuous render for the whole script — chunked renders drift.
• elevenlabs_settings does nothing on cloned voices (HTTP 400, silently retried without). emotion is unavailable too. Speed and pitch are the only real levers.

Raw takes

• iPhone HDR (HLG / Dolby Vision) must be tone-mapped to SDR BT.709 first, or it grades washed out. Check the fps survived: avconvert once turned a 60 fps take into 18.75 fps.
• Use Apple's tone map, never hand-rolled HLG maths (3 Oct 2026). Reel1 on 2 Oct used a numpy HLG curve (gain 0.8) and shipped overexposed and orange: face median luma 148-167 vs Apple's 116, ~10% of pixels clipped, and it was only caught after posting. Extract source frames with tools/hdr_to_sdr_frames.sh raw.mov need.txt src 1350 2400 60 (AVAssetReader → BT.709, VFR-safe, all frames, drop-in for an ffmpeg fps=60 extract). Then put one face frame next to avconvert -p Preset1920x1080 of the raw and look.
• Encode the final as tagged limited-range BT.709, not the yuvj420p that JPEG frames give by default: -vf "scale=in_range=pc:out_range=tv:in_color_matrix=bt601:out_color_matrix=bt709,format=yuv420p,setparams=range=tv:color_primaries=bt709:color_trc=bt709:colorspace=bt709" -color_range tv -color_primaries bt709 -color_trc bt709 -colorspace bt709 -c:v libx264 -profile:v high -preset slow -crf 17 -c:a aac -b:a 320k -ar 48000 -movflags +faststart. tools/final_gates.sh now fails anything else (or < 8 Mbps).
• Bitrate: never below 8 Mbps, aim for ~14. CRF 17 usually lands there on busy frames; when the gate reports less (flat graphics compress hard), swap -crf 17 for -b:v 14M -maxrate 20M -bufsize 28M. Platforms re-encode anyway; a thin upload just gives them less to work with.
• DJI Pocket 3 "glamour" .mov is SDR but tagged FULL range (5 Oct 2026). It needs no tone map, but decode it as full range and output limited-range BT.709, or faces render ~12 levels too dark with crushed blacks: zscale=rangein=full:range=limited where zscale exists, otherwise -color_range pc on the input plus scale=in_range=pc:out_range=tv:in_color_matrix=bt709:out_color_matrix=bt709 (frame extracts: scale=in_range=pc:out_range=pc so the JPEGs keep the real levels). Check the tag first: ffprobe -v error -select_streams v:0 -show_entries stream=color_range,color_transfer -of csv=p=0 raw.mov (pc,bt709 = this case; arib-std-b67 = iPhone HLG, use the Apple tool).
• Skin-luma gate for every raw take: the final's face must sit within ~2 levels of the source at the same moment. tools/skin_luma_check.sh raw.mov <t_src> final.mp4 <t_final> <src_box> <final_box> (boxes are x:y:w:h on a patch of skin; it decodes each file with its own range). A range mistake shows up as ~10+ levels.
• Keep the last take of any repeated line. Cut dead air and false starts. Never cut mid-word.
• Cut private third-party detail he says in passing (a friend's employer, a client, money). List the cut in QA so the owner can restore it.

Silence trimming — the setting that matters most

Reels are hard-edited to almost zero pauses (a reference reel had one 0.10s pause in 39.5s). A raw take has ~26. Trim them — but:

Use a −50 dB threshold. Never −38 dB. On this voice 25% of frames sit below −41 dB, so −38 dB classifies consonants, word tails and breath as silence and cuts speech. That is exactly what "choppy / AI slop" sounds like. Settings: MIN_PAUSE 0.16–0.22, KEEP 0.10–0.13.

Never time-stretch the voice to fit the picture — re-time the picture. Audio stretching is audible at ±10%; video stretching is invisible at ±25%. Cutting silence is inaudible at any amount, because the avatar is motionless during a pause (measured frame delta 0.02/255).

Part 3 — The edit

The three layouts

• FULL — the face fills the frame. Hooks and personal beats.
• GFX — a full-screen graphic, no face: real UI, a stat, a card, a claim.
• PIP / split — graphic on top, face in a card at the bottom (x0 y1087 1080x833 r48). Show a logo or a screen while you keep talking.

Switching every 2–5s is what reads as "edited". Open on FULL or PIP, break to GFX on each key beat, land the CTA on GFX then close on FULL.

The hook: official brand logo (owner rule, 2 Oct 2026)

"The hook can use official brand logos. If the video is talking about Claude, show the Claude logo, for pattern recognition."

• In the hook (first ~2s), when the topic is a named product or brand (Claude, GitHub, Metricool, OpenAI, Blender…), show that brand's official logo.
• Fetch it from the brand's official press / brand page or the official site's own SVG. Never redraw it or AI-generate it.
• Show it as a clean panel element near the caption (the split top panel or a card above the caption), never a pill or sticker.
• Don't alter its colours or proportions. Don't imply endorsement: it names the subject (nominative use), nothing more.
• Cache every logo in brand-logos/<brand>.<svg|png> with a SOURCES.md line per file: brand, source URL, date fetched. Reuse from the cache, and re-check the source if the brand rebrands.

Pipeline order

0. One folder per reel, everything inside it. Each edit (and each parallel editor agent) keeps ALL its work files (frames, panels, audio, renders, logs) inside its own reel folder. Never use a shared _work dir: a parallel run wiped another reel's shared work folder on 5 Oct 2026. Adapting a source reel? Pull its keyframes (a contact sheet every ~0.5s) and use them as the layout and timing reference: which layout when, how long each beat holds, where the text sits. Rebuild it with your own real assets (your take, live screenshots, official logos). Never reuse its footage or graphics.
1. Prep the take (HDR → SDR, fps check) and tighten it (Part 2).
2. Transcribe the tightened audio with Whisper.
3. Author the shot plan — (panel_kind, layout, anchor_phrase, content). anchor_phrase is a phrase from the script, matched against the transcript so the shot starts when those words are actually said. Never hand-type timestamps. Snap boundaries to script words so a cut never lands mid-name ("Matt" / "Pocock's").
4. Capture screens and fetch logos (below). Look at them before using them.
5. Render panels with Remotion, each to its exact shot duration.
6. Composite per shot with ffmpeg, concat, lay the continuous voice over. Run every render under nice -n 15: overload has dropped the creator's mic audio while they were recording. Look-dev on stills first, then do one full render.
7. Burn captions on every word (below).
8. Mix voice + SFX and master (Part 4).
9. Run the QA gates (Part 5). Do not ship if any fails.

Captions — every word, from the first frame

Burned-in captions run continuously from the first word to the last. Punch phrases and kinetic type are extra, never a replacement.

• Caption TEXT comes from the script; only the TIMING comes from Whisper. Using Whisper's text put "THE CARPOTHE" on screen instead of "THE KARPATHY".
• Fix name spellings in captions and the SRT: Claude (not Cloud/Claud), Metricool, GitHub, Fable, Opus, Sonnet, Haiku, and every product named in the reel. Keep a fix map and apply it to every Whisper pass.
• Cue timing comes from a re-whisper of the FINAL audio (turbo), not the raw. Raw-take times shifted by the edit failed the sync gate twice on 2 Oct.
• Shrink long words to fit or "OVERCOMPLICATING" clips to "OVERCOMPLICAT."
• PIP captions need capY≈0.38 — at 0.50 a two-line cue runs under the avatar card and gets clipped. Captions never cover the mouth.
• Caption only the shots you replaced. Kept B-roll already has its own.

Screens — clock-controlled headless capture

Screen-recording a monitor (or filming it with a phone) gives jitter, moiré and a menu bar full of private things. Capture instead:

• Playwright, headless only. Never open a visible browser window.
• Freeze the page clock: an init script replaces performance.now, Date.now and requestAnimationFrame, then step it at 60 Hz and keep every 2nd frame for 30 fps. Every frame is exact, however slow the machine is. Script: tools/capture_clock.js.
• Capture at the reel's size (e.g. 432x768 CSS at 2.5x = 1080x1920).
• Public pages logged out. Recapture live numbers (stars, trending) the day of render and note the capture time in QA.
• Highlight with outline boxes anchored to the real UI element, wiped in on the spoken word. Pan and push-in on the capture; no fake UI. A vendor's claim is shown as their own screenshot with "source: X README", never restated as your graphic.

Framing

Derive crops per take — never reuse them. A crop measured on clone 11 cut clone 13's chin off. Solve for targets:

H = (chin - head) / (chin_pct - head_pct)     y0 = head - head_pct * H
W = H * (out_w / out_h)                       x0 = centre_x - W/2

FULL targets head 13% / chin 85%; PIP 9% / 85%.

Measure the face with YuNet (pipeline/facedet.py, needs opencv-python-headless). Skin-tone thresholds fail silently against a warm wall. YuNet's box is brow-to-chin, so top-of-head ≈ y - 0.55h.

Part 4 — Audio: voice + SFX, no music bed

Raw-voice videos ship voice + synthesised SFX only. No music bed.

• Crisp voice chain: high-pass 70–80 Hz, light denoise (afftdn nr ~8).
• SFX are synthesised, never licensed. Content ID matches fingerprints of known recordings; generated audio has nothing to match. "Royalty-free" libraries are registered in Content ID and CC0 gets fraudulently claimed.
• Few SFX: a sub hit on the hook word, a soft tick or whoosh on graphic cuts, clicks on highlights. 12–20 dB under the voice peak. If you notice them over the words, they're too loud.
• Master to −14 LUFS integrated, ≤ −1.5 dBTP with gain + a true-peak limiter, iterated and measured on the final mp4. loudnorm two-pass failed on peaks on 2 Oct, and its linear mode silently falls back to dynamic.

Part 5 — QA gates (all must pass)

tools/final_gates.sh final.mp4 runs 3, 4 and the contact sheet; tools/caption_sync_gate.py runs 2. Write results to QA.md.

1. Framing — head-top >3% and chin <94% of every crop, measured with a real face detector. Catches the chin-cutoff class of bug.
2. Caption sync, per chapter, on the FINAL render. Re-transcribe the finished mp4 (turbo), match caption words to it, and gate each chapter: median ≤ 0.12s, p95 ≤ 0.25s. Cross-check against the audio's rising-energy onsets. On a fail: retime that chapter's cues to the final words, re-render only the affected frames, re-whisper, re-gate. Loop until every chapter passes (reel4 needed two loops on 2 Oct).
3. Audio length == video length (within one frame), checked with ffprobe right after every render. Overload has silently dropped audio before.
4. Loudness — −14 ±1 LUFS, true peak ≤ −1.5 dBTP. No black frames.
5. Contact sheet — 12 evenly spaced frames in one image. Open it and look at it before saying done. Fail on: any pill, badge or sticker, captions over the face, a wrong or distorted logo, private data, graphics cut by the safe zones or the 4:5 feed crop. Every label and card must sit inside the 4:5 centre crop (1080x1350, y 285-1635), not just the captions.
6. CLAIMS — no number on screen without a source file you can name (path + date). Measured ranges beat spoken roundings ("67–86 sec", not "about a minute"). Anything he says on camera that the source doesn't support stays off screen and is listed under CLAIMS in QA.md for the owner to decide. Verify every on-screen number live (API or the live page, the day of render) and note the capture time; a number from memory or an old note is not a source. A spoken line that is factually false gets cut from the edit, not captioned, and is flagged under CLAIMS with what the source says.
7. First word survives — re-whisper the first 2s after any head trim.
8. Colour — final_gates.sh checks tags and bitrate (yuv420p, tv range, BT.709, >= 8 Mbps); tools/skin_luma_check.sh checks the face is within ~2 levels of the source (Part 2).

Deliverables: final.mp4, contact.png, captions.srt (re-whisper of the final, spellings fixed), QA.md, and a LOG.md written as you go (agents die after 10 min silent).

Gesture-anchored graphics (verdict-card / pointing reels)

Learned on the bad/good/great tools reel (Sep 2026), which shipped mirrored once. When on-screen graphics are things the speaker POINTS at:

• Map the pointing direction before placing anything. Extract a frame at every reveal word and look at where the finger actually is. A right-handed speaker pointing overhead sweeps RIGHT -> MIDDLE -> LEFT from the viewer's side, the mirror of reading order. Never assume left-to-right.
• QA gate: finger matches reveal. After rendering, re-extract a frame at each reveal moment and verify the revealed element is on the side being pointed at. Head-clearance and timing checks do not catch this class.
• Every curiosity element starts hidden. In blur-reveal formats ALL tiles start blurred, each unblurs only at its spoken word (word start minus ~0.12s). A tile that starts sharp kills the hold.
• Face-detector outliers can be hands. A hand raised in front of the face reads as head_top jumping to 20% for one frame. Before shrinking a layout over one outlier measurement, eyeball that frame: hand or head?
• Logo sourcing: official press / brand page or the site's own SVG first (see the hook rule). GitHub org avatars (github.com/<org>.png?size=460) are the org's own mark for repos. Favicon services (google.com/s2/favicons?domain=X&sz=256) are a low-res last resort, never for the hook. Wikimedia thumbs are UA-blocked from scripts. ALWAYS render a labeled contact sheet of every logo and look at it before compositing: it catches wrong brands and broken files in one glance.
• Trim leading dead air by RMS, not by Whisper. Whisper stamps the first word at 0.00 even when the voice starts at 0.50; measure a 20ms RMS envelope, cut to onset minus ~0.1s, then re-transcribe the trimmed file's first 2s to prove the first word survived.

Editorial / "Vox-style" edits (learned Sep 9 2026, one cut rejected first)

Text panels on a flat color are NOT editorial style; they are slides, and the owner reads them as AI slop. Real Vox / Johnny Harris short-form is:

• real imagery cutaways (screenshots, B-roll, the creator's own prior work) with slow push-ins; text sits ON imagery, never alone on a card
• "receipt" cards: source masthead + headline / repo + live star count, dropped over the imagery as proof
• ONE bold sans, white captions, key words in a solid yellow highlighter box; a yellow name lower-third in the first 3s; big stat type over darkened footage
• punch-ins on the talking head on hard beats; calm 8-12 frame ease-outs; numbers count up, highlighter boxes wipe in
• when the speaker points and names a graphic, a graphic MUST be there, anchored to the measured hand position (detect the hand; do not guess)
• build in Remotion (studio/), not Pillow overlays, so everything actually moves
• numbers: one source of truth, verified via API the day of render; QA every frame of a count-up's life, not one sample (a mid-ramp frame reads as a wrong number)

Recurring pitfalls

• Footage↔audio alignment is a QA gate, not an assumption. A Remotion <OffthreadVideo> inside a <Sequence from={X}> restarts at frame 0 unless it has startFrom={X} — every shot after the first then shows footage from the wrong time while the audio is right (lips and hands stop matching). Ship-blocking, and invisible to face/safe-zone/caption gates. Gate: match face-box (or hand) trajectories of the finished file against the raw take at 0.1s across the whole timeline; offset must be 0.0 in every shot.
• Never cut to moving footage of the speaker from another take. Their lips move on different words and it reads as a glitch. Cutaways of prior work are freeze-frames with a push-in, cropped toward the graphics.
• Count-ups must never display a wrong intermediate value where the number is a claim (a frame reading "54% automated"). Hard-in claims; animate only decoration.
• Platform safe zones (shipped cropped once). Instagram crops reels to 4:5 in the feed and lays UI over the edges of the full 9:16 view. Keep ALL critical graphics (cards, labels, text) inside: top >= 220px, bottom <= 1620px, sides >= 60px on a 1080x1920 frame. A card row that starts at y=60 gets its labels cut off on phones. QA gate: check the 4:5 center crop (1080x1350, y 285-1635) still shows every graphic that carries meaning. Captions at y 1612 graze the crop; y ~1560 is safe.
• Forward panel props generically. Copying a hardcoded list of prop names means a new panel kind silently gets undefined and Remotion dies with "outputRange must contain only numbers".
• Never hardcode content inside a panel. Twice a panel shipped showing a previous video's repos and checklist. Panels take data as props, always.
• Sample frames proportionally, not at fixed timestamps — a shorter take makes fixed offsets run past the end and the build crashes.
• Nothing private on screen: no finances, revenue, client names, brand or account ids, or internal files. Tables copied from internal files drop any row with a dollar figure or a client.

Rules that keep the edit from looking like AI slop

• No floating pills, bubbles or tags. Ever. (owner, 28 Sep 2026: "remove the top bubble thing and never have those... looks like AI slop.") Banned:

  • glass or rounded "pill" badges floating over or above the head, e.g. "Opus 5.5" or "Claude Code"
  • a corner tag on every frame, e.g. "AI EDIT", "AI AGENTS" or "PART 1"
  • joke or "random graphic" stickers and chip clusters
  • a REC chip

  Instead, name a tool with its real product UI or official logo as a proper cutaway or split panel: a screenshot on a card, or the actual app window. Otherwise let the caption carry it. Graphics must look like a professional editor made them: fewer, bigger, purposeful, and anchored to a layout (split top panel, full-screen cutaway, lower third). Nothing should float over the face.

  QA gate: in the contact sheet, any small badge, pill or sticker = fail.

• No code-generated motion-graphics sizzle. Use founder footage and real product screens.
• Captions timed by Whisper on the final audio, never guessed. Wrong timing is the #1 tell.
• One accent colour. Serif-italic OR bold-sans captions, not both at once.
• Show, don't narrate: when the voice names a thing, that thing is on screen.
• Use official logos where a brand is named (hook rule above).

Changelog

• 1.2.1 (5 Oct 2026) — DJI Pocket 3 full-range decode + skin-luma gate (tools/skin_luma_check.sh); bitrate floor 8 Mbps, aim ~14; one folder per reel, no shared _work dir for parallel edits; labels and cards inside the 4:5 crop; adapting a source reel = its keyframes as layout/timing reference with your own real assets; verify every on-screen number live; cut and flag factually false lines.
• 1.2.0 (2-3 Oct 2026) — official brand logo in the hook; caption sync gated per chapter on the final render; Apple HDR tone map (tools/hdr_to_sdr_frames.sh) and the colour/encode gate in tools/final_gates.sh.