burn-subtitles

Word-timed subtitle burn-in for shorts. Highlights the current word, no manual sync.

Invoke

.claude/skills/burn-subtitles/burn-subtitles.sh <input> <transcript.json> <out> [style] [font_size]

• input: video path
• transcript: path to JSON. For style=chunks (default) this is a chunks.json from the chunk-captions skill. For other styles it's a transcribe output JSON.
• out: output video path
• style (optional): chunks (default) | line | word-karaoke | selective
• font_size (optional): default 72

Output

mp4 with subtitles burned in. Same dimensions, audio stream-copied. Prints the output path to stdout. Idempotent: skips work if out is newer than both inputs.

Styles

• chunks (default): consumes a chunks.json from the chunk-captions skill. Each chunk is shown as a complete phrase for its [t0, t1] window; the currently-spoken word is rendered cyan, all other words white. Chunks hard-cut between each other — no scrolling, no fade. Kills the per-word slide-up that made the legacy word-karaoke style hard to read.
• line: one plain line per transcript segment.
• word-karaoke (legacy): rolling window of up to 4 words, tied to word timestamps. Each word slides up 8px over 80ms at its own t0 and hard-cuts off when bumped or after a >0.6s speech gap. Deprecated for shorts — feedback showed the constant scroll was hard to follow.
• selective: only burns segments overlapping high-RMS-energy seconds.

Brand preset (deliberate, not default)

This is the project's caption identity — keep it consistent:

• Font: Impact (/System/Library/Fonts/Supplemental/Impact.ttf) — chunky condensed sans, distinct from the CapCut/Submagic/Opus default of Arial.
• Accent (active word): Sapphire Glow #2E6BFF — the C0BALT CUT brand accent (brand/BRAND.md), chosen against the ubiquitous MrBeast-yellow #FFD633.
• Context (earlier visible words): off-white #F5F5F0.
• Position: upper-third (top of first line at ~22% of frame height) — not the default lower-third.
• Animation: 8px slide-up over 80ms (linear) on entry; hard-cut on exit. No ease-out-back pop, no scale animation.
• Stroke: thick black outline (font_size / 10), for legibility over any background without falling back to white-on-black bars.

How

The local ffmpeg build ships without libass, subtitles, or drawtext filters. So burn_subtitles.py renders a transparent PNG per video frame with PIL using the brand preset above, hardlinking duplicate frames for speed. burn-subtitles.sh then composites the sequence onto the source with ffmpeg's overlay filter and re-encodes video (libx264 veryfast crf18); audio is copied.

Status

Works. Tested on a 6.7s 720x1280 clip (say-synthesized speech, 19 words):

• word-karaoke: 20 unique states, frame at t=3.2s shows 16130 white + 5616 accent (highlighted-word) pixels in the lower third.
• line: 27049 white pixels, 0 accent. Dimensions and duration preserved, audio stream-copied (aac).