editorial-short

Editorial short (Fringe Lines)

The look is modeled on Greg Isenberg's shorts. Hold every short to it:

• Never sit on a talking head for more than ~6 s.
• A graphic or layout change lands every 3–5 s.
• Captions appear word by word, with a few orange italic keywords.
• Graphics follow what is being said, so a graphic lands on the word it illustrates.

0. Prerequisites (per machine)

• pipeline/setup.sh installs the venv ~/.shorts-skill and the Remotion node_modules in ~/.cache/fringe-lines-video.
• The episode transcripts must exist: work/<ep>/words-quinn.json and words-doom.json (from pipeline/transcribe_verbatim.py).
• Remotion source: pipeline/remotion/src/ShortV2.tsx (layout) and Graphics.tsx (graphic types).

1. Pick the moment (30–50 s)

• Search the word files for a self-contained take that has:
  • a strong opening line (claim, number, or confession);
  • one idea;
  • a payoff or punchline near the end.
• Good signals:
  • concrete numbers (percentages, dollar amounts, headcount);
  • named companies or products;
  • a contrarian take;
  • a personal experiment ("I let Codex…").
• Avoid clips that depend on earlier context ("like I said", "that thing").
• Several --ranges may be stitched together. Cut at sentence ends, and check the seam doesn't repeat words.

2. Cut it

Add a line to work/<ep>/shorts.sh and run it:

b --name short-NN --range START:END [--range ...] --replace "27=>27%" --replace "claude=>Claude"

• build_short.py does the following:
  • removes fillers, stutters and long pauses;
  • picks the active camera;
  • writes work/<ep>/short-NN/clip.mp4 and props.json;
  • renders the plain v1 short.
• --replace fixes caption text. An empty right-hand side hides that word.
  • Always fix: numbers ("27"→"27%"), product names (Claude, Codex, Descript, GTM, SQL), and Whisper mishearings.
• To get the exact output timings for writing the spec, print them: python3 -c "import json;p=json.load(open('work/<ep>/short-NN/props.json'));print(' '.join(f\"{w['text']}@{w['startMs']/1000:.1f}\" for w in p['words']))"

3. Write the style spec: work/<ep>/short-NN/style.json

{
  "hook":  {"text": "Microsoft wants *27%* of its staff to quit", "end": 2.8},
  "outro": {"text": "Grow *without* growing *headcount*", "hold": 2.0},
  "emphasis": ["microsoft", "27%", "leave"],
  "scenes": [
    {"at": 0.0,  "mode": "full"},
    {"at": 3.0,  "mode": "graphic", "graphic": {"type": "stat", "kicker": "Microsoft buyout offer", "value": 27, "suffix": "%", "label": "of employees offered a payout to *leave*"}},
    {"at": 6.2,  "mode": "card"}
  ]
}

Times are seconds in the cut clip, taken from the word timings in step 2.

Rules that keep it consistent:

• Structure:
  • Scene 0 is always full, with a hook of 4–8 words that ends at ~2.5–3 s.
  • Say the claim; don't summarize the clip.
• Rhythm:
  • Alternate modes as full → graphic → card → graphic → card … with 3–6 graphics in a 40 s short.
  • Use full again for one punchy reaction or laugh in the second half.
• Timing:
  • Start each graphic 0.1–0.3 s before the word it illustrates, and keep it up 2.5–5 s.
  • Two graphics can follow each other directly, e.g. a $20 card then a $100 card.
• Emphasis: pick 10–18 keywords: numbers, names, and the punchline words. Matching is lower-case, ignoring punctuation.
• Accents: in any text field, *word* renders italic orange. Use one or two per string.
• Outro: a 2 s end card with a short line that restates the takeaway, or a question to the viewer.
• Honesty:
  • Numbers on screen must be what was said on the show.
  • Charts are schematic: use lines with shapes, never invented axes or values.
  • Hedge with "roughly" or "~" when the speaker did.

Graphic types (Graphics.tsx):

type	fields	use for
stat	kicker, prefix, value (number, counts up), decimals, suffix, label	one headline number
people	kicker, value (string), label, filled (0–1)	headcount / share of people
lines	kicker, series: [{label, shape: up|flat|down}], caption	trend vs trend
quote	text, by?	the punchline, said on screen
headline	source, title, highlight	news, announcements, a "verdict"
list	title, items[]	steps, products, what went wrong

If a new graphic type is needed, add it to Graphics.tsx and its union type. Keep the card style: white or ink card, 36px radius, a spring entrance, and an animation that runs off its local frame f.

4. Render

~/.shorts-skill/bin/python pipeline/style_short.py --ep <ep> --name short-NN --spec work/<ep>/short-NN/style.json

• The face track (pipeline/face_track.py, YuNet) is computed once and cached in short-NN/faces.json. It keeps the round PiP and card centred on the face; delete the file if the clip is re-cut.
• Sound effects are synthesized: a whoosh on each scene change and a pop when a graphic lands. They are mixed at --sfx-db -20.
• Loudness: −14 LUFS.
• Output: out/shorts/<ep>-short-NN-v2.mp4.

5. QA (always, before reporting)

Pull frames and look at them:

for t in 1.5 5 10 15 20 25 30 35 40; do ffmpeg -v error -y -ss $t -i out/shorts/<ep>-short-NN-v2.mp4 -frames:v 1 -vf scale=270:-1 /tmp/q_$t.png; done

Check each of these:

• Hook: readable, and not covering the face.
• Graphics: they don't overlap the caption block, which sits at y≈1340–1600. No text overflows a card. Long stat values like "+$500K" shrink automatically.
• PiP: the face is centred in the circle.
• Captions: no stray *; keywords render orange.
• Loudness: ffmpeg -i out.mp4 -af ebur128 -f null - 2>&1 | grep " I:" reads about −14.

Report to the user:

• What the short is about and its length.
• Any numbers to verify.
• Any explicit language left in.