caption-burner

---
name: caption-burner
description: Turns a short-video transcript into burned-in captions that hold people to the end, with word-level timing, safe-zone placement and a style spec per platform. Use before publishing any Reel, Short or TikTok.
---

# Captions that keep people watching to the end

A word-level caption file, placement that clears every platform's buttons, and a style card an editor can apply in minutes.

## Claude does

1. Ask for four things and refuse to guess any: the **transcript**, the **platforms** (TikTok, Reels, Shorts), the **spoken language**, and whether the lower third holds faces or UI. No transcript means ask for one, never invent words.
2. Say the reason once: surveying 5,616 US adults, 80% said captions make them likelier to finish a video and 50% usually watch on mute. Write for mute first.
3. Clean the transcript by hand. Around 70% of auto-generated clips need fixing: names, numbers, brand spellings, punctuation. Mark anything unclear as `[unclear]` and ask.
4. Chunk into groups of **1 to 2 lines, 3 to 5 words per line**, one thought each, changing on screen every **2 to 4 seconds**. At roughly 75 spoken words per 30 seconds that is a group every second and a half.
5. Place text inside a **900 x 1400px centred rectangle** on a 1080 x 1920 frame, block between **Y=1200 and Y=1550**. Dead zones to clear, top/bottom/right: TikTok 130/350/164, Reels 220/450/120, Shorts 120/300/48. Call these a working spec to confirm in a live preview, not a platform guarantee.
6. Pick one style for the whole video: **Bold** (all caps, yellow active word, spring pop-in), **Bounce** (70-120-100% scale) or **Clean** (fade-in, white with shadow).
7. Keep the first 1 to 3 seconds to five words or fewer. On Shorts the opening frame is the thumbnail, because almost all views arrive from the feed.

## Then Codex does

Codex takes the bulk arithmetic: approved chunks into a timed subtitle file, every word's start and end computed, one placement variant per platform. Repetitive maths across dozens of cues should not be done by hand.

```
codex exec --sandbox danger-full-access --skip-git-repo-check -C "<working folder>" - < CODEX.md
```

Fill in first: `{TRANSCRIPT}` (cleaned), `{CHUNKS}` (approved groups in order), `{PLATFORMS}`, `{STYLE}`, `{DURATION}` (seconds).

## Claude checks

1. No cue exceeds 2 lines or 5 words per line.
2. Every cue's Y sits inside 1200-1550 and clears that platform's dead zones. A file that renders is not proof it is readable; read the numbers.
3. Caption text matches the cleaned transcript word for word. One added, dropped or improved word is a reject.
4. First cue starts at 0.0s; last cue ends at or before `{DURATION}`.
5. One style only, named in hex values, not adjectives.
6. On any failure, send it back quoting the failing cue numbers. Never patch it yourself or show a half-fixed file.

## Rules

- Auto-captions are a draft, never an output.
- Do not wave off a safe-zone number. Reels raised its bottom dead zone in late 2025 and TikTok ads push it to 370-484px.
- The burned-in caption is not the post caption. Keep them separate.
- Public information only.
- Never invent a fact, a number or a quote.
- Anything sent in someone's name says whose name it is.

## If Codex is not installed

Claude writes the timed file and the per-platform variants itself, and says plainly that it timed them by hand.

## Built from

- social-media-skills/skills, https://github.com/social-media-skills/skills, 66 stars on api.github.com, 5 Sep 2026; the 1-2 line rule, ~70% cleanup, muted-first.
- AgriciDaniel/claude-shorts, https://github.com/AgriciDaniel/claude-shorts, 204 stars on api.github.com, 5 Sep 2026; the three styles, word-level karaoke.
- Verizon Media / Publicis Media study via 3Play Media, https://www.3playmedia.com/blog/verizon-media-and-publicis-media-find-viewers-want-captions/, 5,616 US adults, April 2019; the 80% and 50%.
- sergebulaev/tiktok-skills, https://github.com/sergebulaev/tiktok-skills, 22 stars on api.github.com, 5 Sep 2026; hook-is-not-the-caption.

# Captions that keep people watching to the end - Codex task

You are producing burned-in caption files for a vertical short-form video. You have no memory of any earlier conversation. Work only from what is below.

## You are given

- `{TRANSCRIPT}` - the cleaned spoken words, already human-corrected. Final text.
- `{CHUNKS}` - the approved caption groups, one per line, in order. These are the display units.
- `{PLATFORMS}` - one or more of `tiktok`, `reels`, `shorts`.
- `{STYLE}` - one of `bold`, `bounce`, `clean`.
- `{DURATION}` - clip length in seconds, decimal.

## Produce

1. `captions.srt` - one cue per chunk, numbered from 1, timestamps `HH:MM:SS,mmm`. First cue at `00:00:00,000`, last ends at or before `{DURATION}`. Distribute time by word count across the clip, round to 10ms, no overlaps, no gap over 200ms.
2. `words.json` - an array of `{"word": "...", "start": 0.00, "end": 0.00, "cue": 1}` covering every word of every chunk, in order, for karaoke rendering.
3. `placement-<platform>.json` for each of `{PLATFORMS}`, on a 1080x1920 frame: `{"safe_rect": [90, 260, 990, 1660], "block_y": 1200, "block_y_max": 1550, "dead_zones": {"top": N, "bottom": N, "left": N, "right": N}}` using tiktok 130/350/60/164, reels 220/450/60/120, shorts 120/300/60/48 (top/bottom/left/right).
4. `style.json` - `{STYLE}` as concrete values: weight, capitalisation, hex for base text, active word and shadow, and the animation curve. Use bold = ALL CAPS, base `#FFFFFF`, active `#FFE800`, spring pop-in; bounce = sentence case, base `#FFFFFF`, scale 70-120-100 per cent; clean = sentence case, base `#FFFFFF`, shadow `#000000` at 60 per cent, 120ms fade-in.

## Rules

- Do not change, add, remove or improve a single word of `{TRANSCRIPT}` or `{CHUNKS}`.
- Do not merge or split the supplied chunks. Their count is the cue count.
- If a supplied chunk breaks 2 lines or 5 words per line, output it unchanged and list it under `WARNINGS`.
- Every caption Y stays between 1200 and 1550, inside the 900x1400 centred rectangle.
- One style across the whole video. No per-cue switching.
- Fetch nothing from the internet. Invent no timings for words absent from the transcript.
- If `{DURATION}` is too short for the words at roughly 2.5 words per second, still fit them and flag it.

## Return

Print, in order:

- `FILES:` each filename written, one per line, with byte size.
- `CUES:` total cue count, plus the max words-per-line and max lines found.
- `TIMING:` first cue start, last cue end, `{DURATION}`.
- `PLACEMENT:` per platform, the block Y range and `OK` or `CLASH` against its dead zones.
- `STYLE:` the style name and the hex values used.
- `WARNINGS:` every flag above with its cue number, or `WARNINGS: none`.