Get watched: 10 AI skills for short-form video

hook-lab

write twenty first lines and keep the one that stops the scroll

How the two work together

Claude thinks it through. Paste the Claude prompt into Claude Code, or drop the folder into your skills folder. Claude does the judgement: what to look for, what is worth doing, what is right.

Codex gets it done. At the hand-off point Claude runs Codex on your machine with one command and passes it the Codex prompt. Codex does the mechanical part and hands the result back. Claude checks it before you see it.

No API key to set up: Claude calls the Codex you already have installed. If Codex is not installed, Claude does that half itself and tells you.

Prompt for Claude

---
name: hook-lab
description: Writes twenty opening lines for one short video, scores them, and returns the winner with its on-screen text and opening shot. Use before scripting a Short, Reel or TikTok.
---

# Twenty first lines, and the one that stops the scroll

Twenty scored openers for one video, one winner ready to shoot.

## Claude does

1. Ask for five inputs, refusing to guess any: the **payoff** (what the viewer sees or learns by the end), **audience**, **platform**, **goal** (completion, saves, comments or shares), and the **true public facts**. No payoff, no hooks: the opener proves the promise the video keeps (MrBeast handbook).
2. Set the limits: the spoken hook lands inside 1 to 3 seconds (Hoyos), and at the 2.5 words per second implied by social-media-skills' "~75 words = 30s" that is 7 words on target, 12 maximum.
3. Write the voice ban (sergebulaev/tiktok-skills): no em dashes, no "leverage", "fundamentally", "delve", "harness", "hey guys", no adjective where a number belongs ("3 takes", not "a few takes").
4. Name ten mechanisms: shock, problem-agitation, curiosity gap, story cold-open, social proof, contradiction, hard number, direct callout, before-and-after, question-with-stakes. Strike out any the payoff cannot honestly support, and say which.
5. Save `BRIEF.md` with those inputs, the limits, the ban list and the live mechanisms.
6. Judge the batch, pick the winner, deliver it with the runner-up and why.

## Then Codex does

Codex takes volume and arithmetic: twenty variants across the mechanism grid, plus the mechanical pass on each (word count, banned-word lint, goal tag, five-part score). Bulk work, not judgement. Fill `{PAYOFF}`, `{AUDIENCE}`, `{PLATFORM}`, `{GOAL}`, `{TRUE_FACTS}`, `{BANNED}`, `{MECHANISMS}` into `BRIEF.md`, then run:

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

## Claude checks

1. Count distinct mechanisms. Fewer than eight, reject: twenty variants of one mechanism is one hook.
2. Check every number, name and claim against `{TRUE_FACTS}`. Anything else was invented: delete that hook, do not repair it.
3. Read the winner against the payoff. If the video does not deliver what it promises, reject it whatever it scored.
4. Count the spoken words yourself. Over 12, reject.
5. Grep the banned words yourself. A lint reporting "passed" proves nothing until you have seen it fail.

Any failure: rerun Codex once, naming it in `BRIEF.md`. Twice: Claude writes the rest itself and says so.

## Rules

- Spoken line, on-screen text and opening visual run at once and never repeat each other.
- The caption is not the hook.
- Changing the video idea to earn a stronger hook is the user's call, never silent.
- 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, scores and lints the twenty itself, and says so.

## Built from

- social-media-skills/skills, https://github.com/social-media-skills/skills, 66 stars, api.github.com 5 Sep 2026: scoring process, mechanisms.
- AgriciDaniel/claude-youtube, https://github.com/AgriciDaniel/claude-youtube, 350 stars, api.github.com 5 Sep 2026: variant families, drop-off rating.
- sergebulaev/tiktok-skills, https://github.com/sergebulaev/tiktok-skills, 22 stars, api.github.com 5 Sep 2026: three-track hook, voice ban.
- Jenny Hoyos playbook, https://www.marketingexamined.com/blog/jenny-hoyos-short-form-video-playbook, ~10m views per Short, interviewed by YouTube's Shorts lead: the 1 to 3 second limit.
- MrBeast handbook writeup, https://simonwillison.net/2024/Sep/15/how-to-succeed-in-mrbeast-production/, leaked 36-page internal doc: opener validates the promise.

Prompt for Codex

# Twenty first lines, and the one that stops the scroll - Codex task

You are generating opening lines for one short vertical video. Read `BRIEF.md` in this folder first. You have no memory of the conversation behind it, so `BRIEF.md` is the only truth.

## You are given

In `BRIEF.md`:

- `{PAYOFF}` - what the viewer sees or learns by the end. Every hook points at this.
- `{AUDIENCE}` - who is watching.
- `{PLATFORM}` - Shorts, Reels or TikTok.
- `{GOAL}` - completion, saves, comments or shares. Tag each hook with the goal it serves.
- `{TRUE_FACTS}` - the only numbers, names and claims you may use.
- `{BANNED}` - words and characters forbidden in output.
- `{MECHANISMS}` - the hook mechanisms cleared for this video. Use only these.

## Produce

Write `HOOKS.md` in this folder, containing:

1. A table of exactly **20 hooks**, numbered 1 to 20, across at least **8 mechanisms**, none used more than 3 times. Columns: number, mechanism, goal tag, spoken line, word count, score.
2. A score out of 100 per hook on these weights: hook strength 0.30, clarity within the first second 0.25, emotional intensity 0.20, specificity of the concrete detail 0.15, honesty of the promise against `{PAYOFF}` 0.10. Show the five sub-scores, not only the total.
3. For the **top 3 by score only**, a three-track block:
   - SPOKEN: the line said aloud, 7 words on target, 12 maximum.
   - ON SCREEN: burned-in text, different words from the spoken line, 6 words maximum.
   - OPENING VISUAL: one sentence describing the first frame, shootable on a phone.
4. A lint table per hook: word count, pass or fail on the 12-word ceiling, and every `{BANNED}` term found with its position. Report failures, never silently fix them.

## Rules

- Use only facts from `{TRUE_FACTS}`. Invent no number, name, statistic, quote or result. If a mechanism needs a fact you were not given, write "cannot support: fact not supplied" in that row instead of guessing.
- No em dash characters anywhere. British English.
- Specifics beat adjectives: "3 takes", not "a few takes".
- Spoken line and on-screen text never repeat the same words.
- Hooks only. No caption, no script, no ending.
- Do not edit `BRIEF.md`.

## Return

Print to stdout, in this order, and nothing else:

1. The absolute path of `HOOKS.md`.
2. `HOOKS: <n>`.
3. `MECHANISMS: <n> distinct` and the list used.
4. `LINT FAILURES: <n>` and one line per failure: hook number, reason.
5. `UNSUPPORTED: <n>` and the numbers of rows marked "cannot support".
6. `TOP3: <numbers and totals>`.

If `BRIEF.md` is missing or a placeholder is unfilled, print `BLOCKED: <what is missing>` and stop.

Built from the best public work on this

Sources

  • **social-media-skills/skills** - https://github.com/social-media-skills/skills - 66 stargazers read from api.github.com on 5 Sep 2026 (created 16 Jul 2026, last pushed 3 Sep 2026, actively maintained). 106 skill directories confirmed by listing the contents API. Its `hook-writer` skill supplies the generate-score-shortlist process and the named mechanism list used here.
  • **AgriciDaniel/claude-youtube** - https://github.com/AgriciDaniel/claude-youtube - 350 stargazers read from api.github.com on 5 Sep 2026. A three-layer skill: orchestrator, 14 sub-skills, 9 reference guides. Its `/youtube hook` command produces variants across shock, problem-agitation, story, curiosity gap and social proof, each rated for drop-off risk. Its benchmark percentages are stated in-repo without citation, so they are treated as house numbers and are not repeated in this skill.
  • **sergebulaev/tiktok-skills** - https://github.com/sergebulaev/tiktok-skills - 22 stargazers read from api.github.com on 5 Sep 2026 (last pushed 3 Sep 2026), MIT licensed. Low stars, unusually prescriptive content: its `hook-scripter` writes the first 1 to 3 seconds as three simultaneous tracks, tags 10 hook formulas to a goal, and bans em dashes, "leverage", "fundamentally", "delve", "harness" and "hey guys" while demanding specifics over adjectives.
  • **Jenny Hoyos short-form playbook (Marketing Examined)** - https://www.marketingexamined.com/blog/jenny-hoyos-short-form-video-playbook - practitioner averaging roughly 10 million views per Short with 3 million subscribers, interviewed on YouTube's own blog by Shorts product lead Todd Sherman. Source of the 1 to 3 second hook limit, "you have one second to hook someone", the shock-intrigue-satisfy shape, and the practice of changing the video idea to earn a stronger hook.
  • **MrBeast production handbook writeup (Simon Willison)** - https://simonwillison.net/2024/Sep/15/how-to-succeed-in-mrbeast-production/ - writeup of the 36-page internal document leaked in August 2024 and covered by Tubefilter, Fortune and Cybernews. Source of the rule that the opener must prove the promise the video makes, and that a reveal is placed rather than spent.

Best public prompt we found for this job

**https://github.com/social-media-skills/skills** - the `hook-writer` skill.

Its process, as recorded in our research note when the file was opened on 5 Sep 2026 (a summary of the repo's steps, not a quotation from its text):

"hook-writer gives an explicit process (generate 5-10 candidates by different mechanisms, score, keep 1-2, present 2-3 and recommend one) and 10 named mechanisms"

This skill raises the candidate count from 5 to 10 up to 20, because generation is handed to Codex and volume is cheap there, and keeps the shortlist tight for exactly the reason the original gives: the user is asked to choose between a small number of finished options, not to read a list.

Want this running in your business?

I optimise how businesses run — your sales, your visibility, your social media — and build bespoke software where nothing off the shelf fits. The first conversation is free. Work starts from £150 a day.