Caption styles: 17 looks, built the way the pros set subtitles

Seventeen caption looks with presets you can apply as numbers, not adjectives. Pick how the words are chunked first (a phrase, a group of words or one word at a time), then the look. Every look ships with a plain version that is accessible, since styled word-tracking captions trade some comprehension for attention.

Gates

• Word timings first. Build from word-level timestamps (WhisperX, ElevenLabs Scribe, a forced aligner on the script). Never estimate word times from text length.
• Chunking before look. Phrase captions by default; one word at a time only for hooks, punchlines and clips of 20 s or less.
• Readable on a phone. Keep text inside the 9:16 safe box and at 4.5:1 contrast against what sits behind it; the shared file has the numbers.
• Accessible deliverable. Offer a plain phrase SRT/VTT next to the styled render.
• Free fonts only. Every font in the presets is OFL or Apache. Swap in the person's brand font only if they have a license for video.
• Seek-safe. Every value is a function of timeline time; randomness is seeded by frame. No Date.now(), requestAnimationFrame or unseeded Math.random().
• No credits here. These templates render locally for free. Generated footage, voice or music is the varg-ai skill: call estimate_cost and wait for the person's approval before any generation.

Pick a template

17 templates, 52 presets. Each reference file has the same sections: how the pros do it (every rule cited), presets, timing and numbers, how to build it, mistakes that look amateur, sources. Each JSON file holds the same presets as machine-readable values.

Read references/_shared-caption-craft.md first: it holds the numbers every template here shares.

Workflow

1. Match the request to a template and one or two presets from the table. Say which and why in one line, and show the person the preset names with their one-line looks before building anything.
2. Read the reference for that template, after the shared file, and load its JSON. Take values from the JSON; do not re-guess them.
3. Build it the way the reference's "Build it" section says, starting from the named HyperFrames blocks or varg components where it names them. Drive every value from timeline time so any frame renders the same on a re-render.
4. Fit it to the brand: swap in the person's colors and fonts only where the reference allows, then re-check contrast and safe zones with the numbers in "Timing and numbers".
5. Check a draft: render stills at the key moments (entrance, hold, exit), go through "Mistakes that look amateur", and fix before the final render.

Rules that hold for every template

1. Choose chunking first: phrase (5–12 words), group (2–5), single word only for hooks/punchlines/≤20 s clips (word-for-word: ~90% gaze on text, lower comprehension — Romero-Fresco; Meryl Evans).
2. Active word switches 1–2 frames before audio onset (late noticed from ~45 ms, early unnoticed to ~125 ms — ITU-R BT.1359). Karaoke: line leads 10–20 f (≤1.5 s), fill follows voice (\kf).
3. Per-word motion done within ~6 f at 30 fps; scale via transforms, never font-size.
4. 9:16 safe box: top 288, bottom 672, left 120, right 192 px on 1080x1920; max width 696 px, block centre ~60% height.
5. Contrast ≥4.5:1 against what is behind; black box ≥60% opacity on white footage.
6. Netflix/BBC/DCMP line breaks; 42 cpl 16:9; 2 lines (3 in 9:16); 160–180 wpm; 0.83–7 s per caption; 2-frame gaps.
7. Sentence case; caps only for single-word cards; OFL faces only.
8. Ship a plain phrase version as the accessible deliverable.

Built from

BBC Subtitle Guidelines; Netflix Timed Text Style Guide; Meryl K. Evans (@merylkevans); DCMP Captioning Key; ITU-R BT.1359; Romero-Fresco; WCAG; X: @wayofzorro, @finnardley, @mws, @yt_igm, @tigerjvideo.

Every rule in the references cites its source by number. The full list, with links and what was taken from each, is in references/sources.md. Values marked as house defaults in a reference are our own starting points where no expert published a number.

Related skills

• remotion-captions, iart-caption-animation, cinematic-caption-audrey on varg how to: caption engines these presets can drive.
• short-form-ffmpeg-cookbook: burn SRT/ASS with ffmpeg.
• HyperFrames registry (npx hyperframes catalog --query "…"): about 450 ready-made blocks; the references name the ones to start from.
• varg-video-edit-library: 206 named edits for existing footage (cuts, transitions, headlines, UI, numbers).