short-video-script

---
name: short-video-script
description: Write a 30-60 second TikTok, Reels, Shorts or LinkedIn script with a three-second hook that works muted, one idea, timed lines, on-screen text and shot cues, in two hook variants. Use for any faceless or talking-head short video.
---

# Get a short video watched to the end

Two filmable scripts (hook A, hook B), timed to the second, with on-screen text, shot cues, caption and hook bank, scored before you see them.

## Claude does

1. Collect {TOPIC}, {PLATFORM}, {FORMAT} (faceless or talking-head), {AUDIENCE}, {GOAL} (completion, saves, comments or shares), {PROOF} (a real result or number) and the working folder. Ask once. Never invent a number, result or offer.
2. One idea in one sentence, at most two key points; a third point is the next video, say so.
3. Set {LENGTH}: TikTok 30s, Shorts 45s, Reels and LinkedIn 45s or 60s.
4. Pick the hook formula by {GOAL}: completion, cold-open result; saves, number reveal; comments, open-loop question; shares, relatable call-out.
5. Write {HOOK_A} (stop-scroll: negation, number or question) and {HOOK_B} (retention: story or confession), each opening a loop it does not answer, with 3-7 words of on-screen text in different words and its frame-one visual.
6. Write the stakes sentence with the {PROOF} number.
7. Choose two or three beats in the order that builds, one sentence and one visual change each, then the turn: exception, cost or twist.
8. Choose one {CTA} for the platform: TikTok or Reels "comment [word] to receive" or "link in bio"; Shorts "watch the full video"; LinkedIn "follow for the next one". Decide how the close loops back to frame one.
9. The hook may promise only what a beat shows on screen; otherwise change the hook, never the facts.
10. Assemble {BEAT_PLAN} and write CODEX.md with every placeholder filled.

## Then Codex does

Codex expands the fixed plan into two timed script tables (one per hook), shot cues, cards, word counts, caption, a six-hook bank and the QA score: formatting and counting from decisions already made, not judgement.

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

Fill first: {TOPIC}, {PLATFORM}, {FORMAT}, {AUDIENCE}, {GOAL}, {LENGTH}, {HOOK_A}, {HOOK_B}, {BEAT_PLAN}, {CTA}.

## Claude checks

1. Idea, hooks, beat order and CTA unchanged, and every Rule holds row by row: budget, word count, sentence length, visual changes, card lengths.
2. Read each line aloud; a mid-sentence breath sends it back.
3. Cover the spoken column: the on-screen column alone still tells the story.
4. Bait-and-switch: the result the hook promises is visibly on screen.
5. No em dashes, ALL CAPS, "leverage", "delve", or a vague cue like "[B-ROLL: something relevant]".
6. Eight hooks across at least three archetypes, different angles not rewordings.
7. Rescore the QA checklist. Under 85/100, or any failure above, goes back to Codex with row and reason; rerun until it passes. The user sees only the passing version.

## Rules

- One video, one idea, at most two key points.
- Hook on frame one, done by second 3, showing the result or the tension. No "hey guys", "welcome back", "in this video", logo or slow zoom; the post caption never does the hook's job.
- Hook under 12 words and 50 characters a line, one concrete number or named stake, no vague bait. Never open with "I" without a specific loss.
- On-screen text in different words from the spoken line, never repeated; the video works with the sound off.
- Table: Time | Spoken line | On-screen text | Visual / B-roll.
- Budget, 30s: hook 0-3, stakes 3-10, idea 10-20, turn 20-28, close 28-30; 45s: 3, 10, 35, 42, 45; 60s: 3, 15, 42, 50, 60. Spoken words 65-85, 100-130, 130-160.
- Spoken lines: contractions, short words, no subordinate clauses, 15 words a sentence at most; no ALL CAPS, em dashes or AI words.
- A visual change every 2-3 seconds, never a hold over 5; every cue specific enough to shoot.
- Cards five words or fewer (hook 3-7), sentence case, on the spoken word, inside the safe zone (not the top 20% or bottom 25%).
- One CTA, platform-matched, and the close loops to frame one. Never "like and subscribe", stacked CTAs, "Subscribe!" on Shorts or "follow for more" on TikTok.
- Caption one line, 3-5 hashtags. Vertical 9:16, re-cut per platform; never a reposted horizontal or watermarked clip.
- Deliver at 85/100 or above, then film both hooks and read your own first-3-seconds retention; no source publishes results.

## If Codex is not installed

Claude does the "Then Codex does" section itself and says so.

## Built from

- minhnv0807/ai-business-skills, https://github.com/minhnv0807/ai-business-skills/blob/master/skills/en/04-script-video-global/SKILL.md (570 stars, api.github.com 5 Sep 2026): budget, 50-character lines, 15-word sentences, A/B tables, 85/100 gate.
- alirezarezvani/claude-skills, https://alirezarezvani.github.io/claude-skills/skills/marketing-skill/video-content-strategist/ (25,542 stars, whole repo): 60s structure, captions on, 9:16, repost loss.
- vyralcontent/content-skills, https://github.com/vyralcontent/content-skills (94 stars): three-layer hook, 6-10 hooks, loop-back, safe zones.
- rediumvex/viral-hooks-skill, https://github.com/rediumvex/viral-hooks-skill (81 stars): 12-word hooks, stop-scroll plus retention pair.
- sergebulaev/tiktok-skills, https://github.com/sergebulaev/tiktok-skills (21 stars): formulas by goal, different words on screen, humaniser.
- xnomad-dev/claude-skills, https://github.com/xnomad-dev/claude-skills/blob/main/video-script-writer/SKILL.md (0 stars, craft only): words a minute, cue format.

# Get a short video watched to the end - Codex task

## You are given

- {TOPIC}, {PLATFORM} (TikTok, Reels, Shorts or LinkedIn), {FORMAT} (faceless or talking-head), {AUDIENCE}, {GOAL} (completion, saves, comments or shares), {LENGTH} (30, 45 or 60 seconds).
- {HOOK_A} and {HOOK_B}: finished hooks, each with spoken line, 3-7 word on-screen text and frame-one visual.
- {BEAT_PLAN}: the one idea, the stakes sentence, two or three beats each with a visual change, the turn, the loop back to frame one.
- {CTA}: the single closing line.

Do not change the idea, hooks, beat order or CTA. Expand the plan; do not rethink it.

## Produce

Three files in the working folder: `script-A.md`, `script-B.md` (same plan, hook A or hook B in row one) and `hooks-and-caption.md`.

Each script file:

1. A table, one row per beat: `Time | Spoken line | On-screen text | Visual / B-roll`, timed to the budget for {LENGTH}. 30s: hook 0-3, stakes 3-10, idea 10-20, turn 20-28, close 28-30. 45s: 0-3, 3-10, 10-35, 35-42, 42-45. 60s: 0-3, 3-15, 15-42, 42-50, 50-60. Split a beat into two rows whenever a picture would hold over 3 seconds, so the visual changes every 2-3 seconds and never holds over 5. The close row carries {CTA} and the loop-back visual.
2. Visual column as shootable cues: `[B-ROLL: exact shot]`, `[CUT TO: ...]`, `[TEXT ON SCREEN: ...]`, and `[CAPTION: ...]` on every stat, punchline and the CTA. Faceless: name the shot or screen capture. Talking-head: when to cut away and to what.
3. Spoken word count, written in the file. Budget: 30s 65-85, 45s 100-130, 60s 130-160. Outside it, tighten and recount before saving.
4. Longest sentence, in words.

`hooks-and-caption.md`:

5. Caption: one line for {PLATFORM} plus 3-5 specific hashtags.
6. Hook bank: six more hooks, labelled by archetype (result, number, question, story, confession, call-out, contrarian) and goal, covering at least three archetypes, each a different angle rather than a rewording, each with its own 3-7 word on-screen text.
7. QA score for script A, ten checks at 10 points: hook by 3s; hook 50 characters or fewer a line; one idea; at most two points; spoken tone; sentences 15 words or fewer; one specific CTA; length fits the platform; does not open with "I"; every cue shootable. Score honestly.

## Rules

- Spoken lines: contractions, short words, no subordinate clauses, 15 words a sentence at most. No ALL CAPS, em dashes, "leverage", "delve", "harness", "fundamentally" or "hey guys".
- Cards: five words or fewer (hook card 3-7), sentence case, different words from the spoken line, on the spoken word, inside the safe zone (not the top 20% or bottom 25% of a 9:16 frame).
- Hooks under 12 words, never answering their own question.
- The on-screen column alone must tell the story with the sound off.
- One CTA only, exactly {CTA}; never add "like and subscribe" or a second ask.
- Every cue names a real shot; never "[B-ROLL: something relevant]".
- Nothing in {BEAT_PLAN} is dropped silently: if it does not fit, cut words, not beats, and report it.

## Return

Print at the end:

- Paths of the three files.
- Word count and longest sentence for A and B.
- Rows per table and the longest gap between visual changes, in seconds.
- Row one and the close row of script A, verbatim.
- QA score, naming any check under 10.
- Anything from the plan that did not fit.