reverse-engineer

/reverse-engineer: learn an editing style, ship skill iteration 1

This is how fireship-style-edit was born (2026-09-29: Fireship's "Meta is pivoting again" was measured shot by shot, then turned into an engine + style bible). This skill makes that repeatable.

Output: a study folder (numbers + sheets) and a new skill folder .claude/skills/<style>-edit/ that renders a demo in that style. Learn the grammar; never copy the reference's footage, music, logo or show name.

Step 1. Measure (scripts/study.py)

python3 "${CLAUDE_SKILL_DIR}/scripts/study.py" --url "<reference url>" --slug <style> \
    [--start 0 --dur 300] [--prompt "names, products"] [--out studies/<style>]

Downloads with yt-dlp (runs yt-dlp -U on a YouTube 403), trims to the span you want (default whole video, max 10 min; 2 to 5 min is plenty for a long-form style, the whole thing for a short), then measures:

What	How	Why it matters
cuts, shot length median / mean / p10 / p90, cuts per min, % < 1 s / > 3 s	ffmpeg scene detection (--threshold 0.30, lower = more cuts)	the pacing rule
visual events (things landing: stickers, cards, cutouts) per min + median gap	frame-difference spikes at 10 fps	layered styles (Fireship) change every ~1 s WITHOUT cutting; cuts alone read them as slow
longest shots	from the cut list	the "builds" (diagrams, read-alongs) worth copying
motion, near-static stretches	frame differencing at 4 fps	"nothing static > X s"
face on screen %, face centre, face height	OpenCV face detection	faceless vs talking head vs split / band layout
palette, brightness	k-means on sampled frames	colour tokens
loudness (LUFS), loudness range, pauses	ebur128 + silencedetect	mix + jump-cut density
words, wpm, transcript	Whisper small.en	narration pace, script formula

Writes report.md / report.json (with measured verdicts) and the sheets:

• frames/shots_NN.jpg: ONE frame per shot, numbered, timestamped, with shot length.
• frames/hook.jpg: the first 3 s at 10 fps (what stops the scroll).
• frames/dense_00_open.jpg: the first 30 s at 2 fps (element changes inside shots).
• frames/dense_0N_long.jpg: the longest shots at 2 fps (how builds unfold).

If scene detection misses cuts (fades, whip pans, cuts between near-black cards) or double-counts (flashes), rerun with --threshold 0.15 / 0.3 and compare against the sheets. A cut list you didn't check is a guess. Learned on the first test (Fireship, 2026-10-02): at threshold 0.3 it read 3.65 s per shot because most changes are elements landing on a held black ground, not hard cuts; the visual-events metric (a new element every ~1 s) is what captures that style.

Step 2. LOOK (the part a script can't do)

Read every sheet with the Read tool (they are images). For each recurring element, write down: what it is (card, sticker, meme, chart, cutout, caption, lower third, zoom, transition), how it enters (pop, slide, whip, bulge), how long it holds, where it sits, what spoken moment triggers it, its text style (font family look, colour, stroke, size, case), and how often it recurs. Also: the hook (hook.jpg), the ending, caption track yes/no, layout (full-bleed, split, band, PiP), transitions per minute. Zoom in on anything unclear: ffmpeg -ss <t> -i source.mp4 -frames:v 1 x.png and Read it at full size. Fill the "Fill in from the sheets" block at the bottom of report.md.

Step 3. Scaffold iteration 1 (scripts/scaffold_skill.py)

python3 "${CLAUDE_SKILL_DIR}/scripts/scaffold_skill.py" --study <study folder> \
    --name <style>-edit --title "<Creator / format name>" --base fireship|talking-head|longform|none

Pick the base engine closest to the style: fireship = faceless B-roll over a VO (16:9, or 9:16 when the study is vertical); talking-head = vertical face-in-band compositor; longform = 16:9 talking head; none = write a new engine. It creates SKILL.md with the numbers table and the rules that follow from them, references/study.md, the base templates and a starter render_template.py.

Step 4. Write the style bible (fill every TODO in the new SKILL.md)

• description: one sentence of trigger criteria (when Kevin should get this skill).
• Structure: the section formula with timings from the transcript (hook, setup, beats, payoff, CTA).
• Devices table: one row per recurring element from step 2, mapped to an engine function. If the base engine lacks a device, write it now (deterministic in t, PIL, cached) in templates/.
• Text + captions: closest open fonts (copy the TTFs into templates/fonts/), colours from the palette, stroke, size, position.
• Don'ts: what would break the style, phrased from the numbers. House rules carry over to every new skill: literal visuals (fireship-style-edit section 0), real tweet cards, no em/en dashes in on-screen text, the 10 s third-party clip rule, masters to ~/Downloads.

Step 5. Prove it: demo render

Build a 20 to 40 s demo in the new style using the new skill only (any VO Kevin supplied, or the study's own audio for a side-by-side). Then compare against the reference:

1. Render, then run study.py --file renders/demo.mp4 --slug <style>-demo on your own output.
2. Put the two report.md number tables side by side: shot median, cuts/min, static stretches, face %, wpm, LUFS. Anything off by more than ~25% is a rule the skill didn't encode yet; fix the skill, not just the demo.
3. Tile both hook.jpg sheets next to each other and LOOK.
4. Copy the demo to ~/Downloads/<Style> Demo.mp4 and add a Learnings line to the new SKILL.md.

That is iteration 1. Kevin's notes on the demo drive iteration 2 (edit the SKILL.md rules and engine, same as the Fireship skill's 10-02 literal-visuals pass).

Don'ts

• No skill from numbers alone: step 2 (looking at every sheet) is mandatory.
• No reference assets in the new skill (their footage, music bed, logo, fonts-as-branding, show name).
• No "iteration 1" without a demo render and the side-by-side numbers check.
• Don't overwrite an existing skill (--force only when Kevin asks to redo it).