engaging-monologue

Engaging monologue (16:9)

Reference: Ali Abdaal, "How I'd Create Content in 2026 (If I Had To Start Over)" (youtu.be/A8PxRHQrLJ8, 16 min, 1080p). The motion language is copied from that video. The look is the FringeLines v2.1 kit: My Drive/FringeLines Design System files/Fringe Lines Design system v2_1/ and the tokens in …/handoff/fringelines-tokens.css.

What makes the reference work

Measured from the video at 1 frame per 2 s, plus dense samples around transitions.

• About 55% plain talking head. One locked-off camera, mid-shot. Jump cuts every ~5 s (median): about 122 cuts in 16 min. Each cut alternates wide and punched-in, so the cuts read as intentional.
• Two layouts, never more:
  1. Full-frame speaker with things laid over it: name lower thirds, a coloured callout card top-left with a tiny label pill, white serif sentences at the bottom, small coloured chips near the head, and a sponsor/CTA pill at the bottom centre.
  2. Paper canvas. The speaker shrinks into a portrait PiP (rounded corners, white frame, right side, about 400×560) while a graphic builds on a warm off-white page with one faint swoosh. The canvas carries kinetic sentences, numbered roadmap cards, pill diagrams, pyramids, and screenshots or b-roll in rounded cards with a caption underneath.
• No running captions. On-screen text is chosen. One sentence at a time, revealed word by word in time with the speech. Each word starts faint and blurred, then sharpens. The key phrase changes colour.
• Diagrams build piece by piece as each item is said. Pills pop in with a little overshoot. Pyramid layers drop in from above.
• The camera moves inside graphics. The 5-card roadmap returns at each chapter, zooms into the active card and blurs the others.
• Exits blur out rather than cut. The full↔PiP change is a quick eased morph of about 0.4 s.
• Rhythm: a new visual idea every 10–20 s, one emphasis moment (black-and-white) per section, and a lower third the first time each person appears.
• Sound: soft whooshes on layout changes and light pops when items land, kept well under the voice.

How it maps onto FringeLines (don't drift from this)

Reference	FringeLines version
Warm serif with coral highlights	Instrument Serif for all display text. The highlight is an italic phrase, ink-blue #1A4F6B on paper or white on video. Orange #D4622A only for the one number that matters ([7 AM], the stat).
Rounded sans in pills	Archivo 500–600 for pills, chips, list items, stack labels
Small caps labels	IBM Plex Mono 600, uppercase, +0.14em, ≥ 22 px: eyebrows, label pills, sources, meta
Coral/red gradient callouts	Kit gradients: green (default), blue (second idea), orange (warning or "mistake")
Pastel roadmap cards	Green, blue and orange washes, a mono number badge and a dashed orange connector
Cream paper + swoosh	Paper #F7F7F5 with one faint green-wash swoosh
Drop shadows	None. Flat surfaces, 2 px hairlines #E3E5DF, a white frame on the PiP, radius 16–18
Sponsor pill	White CTA pill (fringelines.co / newsletter) with an orange mono "Link in description" tab
Channel outro	Green end card, the Clip Slides cover: mono eyebrow, orange rule, serif headline with one italic phrase, Archivo sub, ● fringelines.co

The kit rules still apply:

• one idea per screen;
• one orange per screen;
• green is the default, blue the comparison;
• every stat has a source;
• nothing below 22 px.

Pieces

• Template: pipeline/remotion/src/Monologue.tsx, the Monologue composition at 1920×1080 / 30 fps. It reuses the face tracking from ShortV2.tsx.
• Build: pipeline/build_monologue.py, which has two steps, cut and render.
• Work folder: work/monologue/<slug>/ holds clip.mp4, props.json, transcript.txt, beats.json and faces.json.
• Output: out/monologue/<slug>.mp4, normalised to −14 LUFS.
• Worked example: work/monologue/morning-loop-test/beats.json, a 62 s clip of Quinn from the Oct 30 episode.

1. Get a take

A single-camera 16:9 recording with good audio, one person, roughly 1–15 min. Put it in inbox/monologue/<slug>/take.mp4.

To pull a solo stretch from an episode, mux the camera with its studio-sound track. Take the offsets from the Descript XML with build_short.xml_offsets; see how morning-loop-test was made:

ffmpeg -ss <t0-videoOffset> -t <dur> -i Quinn.mp4 -ss <t0-audioOffset> -t <dur> -i studio-sound-Quinn.m4a -map 0:v -map 1:a … take.mp4

2. Cut

~/.shorts-skill/bin/python pipeline/build_monologue.py cut --slug <slug> --src inbox/monologue/<slug>/take.mp4 \
    [--range 12.0:95.5 --range 120:180] [--replace "gtm=>GTM" --replace "quick=>Quick"]

What the cut step does:

1. Transcribes the take with Whisper (cached in words-src.json).
2. Removes fillers, stutters and long pauses, using the same planner as the shorts.
3. Alternates the punch-in (1.0 / 1.12, centred on the face) at each jump cut.
4. Writes transcript.txt, with word@seconds on the cut clock.

--replace only fixes the words used to time the reveals; nothing is captioned automatically.

3. Write work/monologue/<slug>/beats.json

Read transcript.txt and plan the beats against it.

{"brand": "FringeLines", "beats": [
  {"type": "title",   "at": 0.2,  "end": 2.0,  "eyebrow": "FringeLines · Monologue", "text": "The *morning loop*"},
  {"type": "name",    "at": 2.1,  "end": 5.4,  "text": "Quinn", "role": "Host · Fringe Lines"},
  {"type": "chips",   "at": 3.0,  "end": 7.4,  "items": [{"text": "GTM loops", "at": 3.1}, {"text": "Coding loops", "at": 5.1, "tone": "blue"}]},
  {"type": "text",    "at": 7.6,  "end": 11.3, "eyebrow": "What's a loop?", "text": "Loop's just a *stated goal* and then it works *autonomously*"},
  {"type": "callout", "at": 11.5, "end": 16.2, "label": "The setup", "tone": "green", "text": "I use Quick Desktop with routines that execute at any time"},
  {"type": "overlay", "at": 16.4, "end": 18.3, "text": "I usually run it at [7 AM]"},
  {"type": "stack",   "at": 22.4, "end": 36.4, "title": "My core team · 4 people", "items": [{"text": "Account executive", "at": 25.6}, "…"]},
  {"type": "stat",    "at": 51.2, "end": 55.1, "kicker": "My regular account team", "value": 14, "text": "people I'm interacting with on a regular basis", "source": "Quinn's own count · Fringe Lines, Oct 30 2026"},
  {"type": "end",     "at": 58.4, "end": 61.4, "text": "Building the *morning loop*", "sub": "The full walkthrough is in this week's episode."}
]}

Times

• Times are seconds on the cut clip. end (or hold) closes a beat, and its exit blurs out over 0.3 s.
• Text is revealed word by word in sync with the speech. Every word in text is matched to the transcript inside the beat's window.
  • Quote the speaker verbatim (light trims are fine) so the sync works.
  • If the text is a paraphrase, the reveal falls back to an even spread. Force that with "reveal": "even".
• Item at times default to an even stagger. Set them to the moment each item is said, about 0.1 s before the word.

Markup

• *italic highlight*: one phrase per screen.
• [accent number]: orange, one per screen, numbers only.

Beat types

Speaker full-frame (overlays):

type	fields	use
title	text, eyebrow?	Cold-open title over a dimmed frame (first 2–3 s)
name	text, role?	Lower third the first time a person appears
overlay	text	One spoken sentence in white serif at the bottom: the line you want remembered
callout	text, label?, tone?, side?	Coloured card top-left (or right) with a mono label pill: "Mistake #1", "The setup", "Rule"
chips	items[{text, at, tone?}], side?, tone?	2–3 short tags popping in beside the head: inner voice, names, options
cta	text, sub?	Bottom pill: the newsletter, the episode or the offer. Once, usually mid-video
bw	(none)	Black-and-white emphasis on the speaker; pair it with an overlay
media + "full": true	src, video?, from?	Full-bleed b-roll with the speaker in the PiP

Paper canvas (the speaker moves to the PiP):

type	fields	use
text	text, eyebrow?, align?	A kinetic sentence: definitions, the thesis, the "single most important…" line
stack	items[{text, at, tone?}] (bottom→top), title?	Pyramid or layers: foundations → strategy → payoff, team layers
compare	title?, left{label, items}, right{label, items}, winner?	A vs B: hobby vs revenue, before vs after. The loser greys out near the end
roadmap	items[{text}] (≤6), active?, zoom?	The video's chapters. Show once at the start, then reuse it per chapter with active = the chapter index and zoom = when to zoom in
steps	items[{text, at}], title?	A numbered list with hairline rules: specialists, tools, steps
stat	value, prefix?, suffix?, decimals?, kicker?, text, source	One number counting up in orange; source required
media	src, title?, meta?, video?, from?	A screenshot or b-roll in a rounded card with a caption (title + mono meta line). Files go in the work folder
end	text, eyebrow?, sub?, url?, tone?	The green end card (2–4 s)

Paper beats less than 0.7 s apart stay in the PiP layout, so the speaker doesn't bounce.

Pacing rules (from the reference)

• Open with a title and a name, both inside the first 5 s.
• Paper beats take 35–45% of the runtime. Never go more than ~20 s without a visual change, and never stay on paper longer than ~15 s.
• Alternate the two layouts. Don't run two callouts back to back; put a text or a diagram between them.
• One idea per beat. Keep text beats to ≤ 14 words and callouts to ≤ 16.
• Diagram items land on the word, never all at once. A diagram's first item should land within ~1 s of the beat starting; otherwise the beat starts later.
• Sections:
  • Show a roadmap at the end of the intro.
  • Reuse it at each section start with active set.
  • Close each section with an overlay of its key line, with bw for the biggest one.
• One cta per 5 min at most, never in the first 30 s.
• End on the end card. It can run past the last word.

Honesty

• Numbers on screen must be what was said, or come from a named source in source.
• Screenshots of other people's content need credit in meta.
• Don't put words in the speaker's mouth: an overlay or text beat that isn't a quote should read as a label, not as speech.

4. Render

~/.shorts-skill/bin/python pipeline/build_monologue.py render --slug <slug> [--sfx-db -22]

• It tracks the face once (faces.json; delete it if you re-cut).
• Rendering time grows with the video length; a 1-min test took a few minutes on the Mac mini.
• Sound effects: a whoosh on each layout change, a pop when an item lands, a riser into the title and end cards. Raise --sfx-db toward −18 for more energy.

5. QA (always)

for t in 1.5 5 10 20 30 40 50 60; do ffmpeg -v error -y -ss $t -i out/monologue/<slug>.mp4 -frames:v 1 -vf scale=640:-1 /tmp/m_$t.png; done

Check each of these:

• Text: no text runs under the PiP (the paper area ends at x = 1320). Stack labels fit on one line.
• Highlights: no stray * or [. There is at most one orange per screen.
• Face framing: the face is centred in the PiP, and the punch-ins don't cut off the head.
• Reveal timing: reveals land on the spoken word. If a beat shows faint words for a long time, its at is too early.
• Loudness: ffmpeg -i out.mp4 -af ebur128 -f null - 2>&1 | grep " I:" reads about −14.

Report to the user:

• the runtime and the beat list (time → type → text);
• what still needs checking: numbers, names and screenshot credits;
• any beat that fell back to an even reveal, because its text wasn't verbatim.

Not built yet

• A 9:16 version: the composition is 16:9 only. For vertical, cut a short with the editorial-short skill.
• Auto-planning: writing beats.json is a judgment call, so it's done by hand or by Claude from transcript.txt.