script-45s

---
name: script-45s
description: Writes a 45-second short-form script that lands with the sound off: retention structure, muted-first captions, timed shot list. Use for any Reels, Shorts or TikTok script.
---

# A 45-second script that lands with the sound off

One timed script, readable on mute, a shot and on-screen line per beat.

## Claude does

1. Ask five things: the ONE idea, the audience, the outcome for the viewer, the proof you own (number, demo, before/after), the platform. Never invent the proof. No proof, no script.
2. Set length. 45s is the brief; 50-60s is the reach sweet spot in Galloway's 5,400-Short study. Budget ~75 spoken words per 30s, so 110-120.
3. Write the hook as three tracks in seconds 0-3: spoken line, on-screen text, opening visual. Hoyos caps it at 3 seconds. Read it muted; if the text alone does not hold, rewrite.
4. Structure: hook 0-3s, foreshadow the conflict 3-8s, 3-4 delivery beats with a but/then turn between each 8-38s, payoff 38-45s. Every second justifies itself. Place the reveal, never spend it. Emotion peaks last.
5. Mark a visual change every 2-4 seconds and write its shot. Score on the claude-shorts rubric: hook 0.30, coherence 0.25, emotional intensity 0.20, value density 0.15, payoff 0.10. Under 60, rewrite first.
6. Strip banned voice: no em dashes, no leverage, fundamentally, delve, harness, "hey guys".
7. Fill {SCRIPT_JSON}, {PLATFORM}, {CAPTION_STYLE}, then hand over.

## Then Codex does

Codex builds out the approved script: two-column AV table, timecode grid, caption chunking into two-line blocks inside a 900x1400px centred safe rectangle, the other two platform variants with encode specs, three alternative hooks. Repetition across platforms, not judgement.

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

Claude fills first: {SCRIPT_JSON} (approved beats, timings, VO), {PLATFORM} (primary target), {CAPTION_STYLE} (Bold, Bounce or Clean).

## Claude checks

1. Read it with the VO column covered. If the on-screen text alone does not carry the idea, reject. Roughly 80-85% watch on mute.
2. Count words against runtime, ~75 per 30s. Over by more than 10%, rerun.
3. Visual changes land 2-4 seconds apart, no gap over 5 seconds.
4. Trace every number and quote to the proof input. One unsourced figure, reject: the claude-youtube benchmarks are that repo's uncited house numbers, never script claims.
5. No caption above y=120px, below y=1550px, or in the right-hand 164px TikTok action rail.
6. Reject-and-rerun: send back the failing line and the rule it broke. Never patch silently.

## Rules

- Muted first. Write the on-screen text before the voiceover.
- Never present a source's uncited house number as verified data.
- Length is a lever, not a promise. Never sell runtime as a result.
- 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 does the Codex half itself, same formats, and says so.

## Built from

- social-media-skills/skills, https://github.com/social-media-skills/skills, 66 stars via api.github.com, 5 Sep 2026.
- AgriciDaniel/claude-shorts, https://github.com/AgriciDaniel/claude-shorts, 204 stars via api.github.com, 5 Sep 2026.
- Jenny Hoyos playbook, https://www.marketingexamined.com/blog/jenny-hoyos-short-form-video-playbook, ~10M views per Short, interviewed by YouTube's Shorts lead.
- MrBeast production handbook, https://simonwillison.net/2024/Sep/15/how-to-succeed-in-mrbeast-production/, leaked 36-page internal document (Fortune).
- Verizon/Publicis captions study, https://www.3playmedia.com/blog/verizon-media-and-publicis-media-find-viewers-want-captions/, 5,616 US adults surveyed.

# A 45-second script that lands with the sound off - Codex task

Format an already-approved short-form script. Do not rewrite the idea or add facts.

## You are given

- **{SCRIPT_JSON}** - ordered beats, each with `start`/`end` in seconds, `vo` (spoken line), `text` (on-screen words), `shot` (what is on camera). Runtime 45s unless stated.
- **{PLATFORM}** - primary target: `shorts`, `reels` or `tiktok`.
- **{CAPTION_STYLE}** - `Bold` (ALL CAPS, yellow highlight on the key word, spring pop-in), `Bounce` (70-120-100% scale) or `Clean` (white fade-in with drop shadow).

## Produce

Four files in the working folder:

1. **`script-timed.md`** - two-column AV table, one row per beat: `VIDEO` (the shot, tagged `[Action]`) and `AUDIO` (tagged `[VO]`, `[SFX]`), plus a `TC` column, `mm:ss.s - mm:ss.s`. Underneath, the spoken word count and words-per-30-seconds.
2. **`captions.md`** - each on-screen block chunked to 2 lines and 6 words per line max, with in/out timecodes, an X/Y anchor on a 1080x1920 frame, and the {CAPTION_STYLE} name. Every anchor sits inside a 900x1400 centred rectangle, block between Y=1200 and Y=1550.
3. **`shotlist.md`** - one row per visual change: shot number, timecode, framing (wide, medium, close, insert), location, A-roll or B-roll. Then a `BATCH BY` heading grouping rows by location and framing.
4. **`variants.md`** - the script re-cut for the other two platforms. Per platform: safe-zone Y range (Shorts top 120px, bottom 300px, right 48px; Reels top 108-220px, bottom 310-450px, right 120px; TikTok top 130px, bottom 320-350px, right 120-164px), encode spec (all 1080x1920; Shorts H.264 High 4.2, 12 Mbps, AAC 192k; TikTok CRF 18, AAC 128k; Reels 4.5 Mbps max, AAC 128k), and any beat whose text must move to clear that rail. Then three alternative hooks for seconds 0-3, as three tracks.

## Rules

- Change no wording in `vo` or `text` except to chunk it. If a line will not fit, flag it, never reword.
- Invent no numbers, statistics, quotes or claims. Carry every figure through exactly.
- Never let a gap between visual changes exceed 5 seconds.
- No em dash characters anywhere in the output.
- Banned in any new text: leverage, fundamentally, delve, harness, "hey guys". A number beats an adjective.
- British English.
- If a placeholder is still literally `{SCRIPT_JSON}`, `{PLATFORM}` or `{CAPTION_STYLE}`, stop and print `MISSING INPUT: <name>`. Never guess.

## Return

Print to stdout, in order:

1. `FILES:` the four filenames.
2. `WORDS:` spoken word count, `WPM30:` words per 30 seconds.
3. `RUNTIME:` the last beat's end time.
4. `MAX_GAP:` longest gap between visual changes.
5. `CAPTION_BOUNDS:` `OK` if every anchor is inside the rectangle, else each offending beat with its X/Y.
6. `UNSOURCED:` every number, statistic or quote in the output, or `NONE`.