carousel-builder

---
name: carousel-builder
description: Turns one idea into a slide-by-slide carousel people swipe to the end of. Use for a LinkedIn or Instagram carousel that gets saved and shared.
---

# A carousel people swipe to the end of

You get the full slide script, hook to close, already written.

## Claude does

1. Ask for four things, guessing none: **{TOPIC}**, **{AUDIENCE}**, **{PLATFORM}** (LinkedIn or Instagram), **{PROOF}** (the real number, story or example you own). No proof, no carousel: say so and stop.
2. Read `about-me.md` / `voice.md` if they exist. Both leading repos draft from a persistent voice layer and ask only for gaps. Ask once.
3. Pick one of Haines' five architectures and name it: **Listicle, Story arc, Before/After, Framework/Steps, Myth-bust**. Choose from the topic, no menu.
4. Set slides to **6-10**. Fewer has no swipe momentum, more loses the tail.
5. Slide 1 is a hook of **40 characters or fewer**, readable at thumbnail size. It states the payoff, never teases it.
6. Middle slides carry **one idea each**: headline **50 characters or fewer**, body **25 words or fewer**.
7. End every slide on an open loop the next closes: an incomplete count, an unexplained term, a promised number.
8. Give **{PROOF}** its own slide two thirds through, source named on it.
9. Final slide asks **one action**. Instagram: comment CTAs measure **+202.78% comments**, questions **+36.70%**.
10. Caption is separate: **200-250 words**, hook inside the first 125 characters, **no hashtags** (**-31.70% views**).
11. Score against the user's own top 10% if they have an export, engagement = reactions + (comments x 3). Generic best practice is the fallback.

## Then Codex does

Codex takes the approved plan and does the bulk: expanding each slide to final text, enforcing every character and word limit, writing three hook variants, producing the handoff sheet. Repetitive counting at volume, done faster.

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

Fill in first: **{TOPIC}**, **{AUDIENCE}**, **{PLATFORM}**, **{PROOF}**, **{FRAMEWORK}**, **{SLIDE_COUNT}**, **{VOICE_NOTES}**.

## Claude checks

1. Slide 1 over 40 characters: rerun.
2. Any headline over 50 characters: rerun.
3. Slide count outside 6-10: rerun.
4. Any middle slide carrying two ideas: rerun.
5. Any hashtag in the caption: delete it. Prompt libraries still assume hashtags help; a 24.36M-post study says they cost views.
6. Any statistic not in {PROOF} or the sources below: cut it. The circulating "2.5x more shares" and "+27% completion" claims trace to no measurement firm.
7. Final slide asking more than one thing: rerun.

Quote the failing count back and rerun. Never hand-patch.

## Rules

- Public information only.
- Never invent a fact, a number or a quote.
- Anything sent in someone's name says whose name it is.
- No hashtags.
- Build for carousels on purpose: LinkedIn median engagement 21.77% against 7.35% for video, 9x the saves of a single Instagram image.
- Never repeat a precise-sounding statistic with no stated sample behind it.

## If Codex is not installed

Claude writes the slide file and handoff sheet itself and says it did that half by hand.

## Built from

- charlie947/social-media-skills, https://github.com/charlie947/social-media-skills, 3,283 stars via api.github.com; gemini-carousel and post-scorer.
- coreyhaines31/marketingskills, https://github.com/coreyhaines31/marketingskills, 46,945 stars via api.github.com; the five carousel architectures.
- Metricool 2026 Instagram Study, https://metricool.com/press-release-instagram-study-2026/, 24,364,803 posts; hashtag, saves, CTA figures.
- Metricool 2026 LinkedIn Study, https://metricool.com/press-release-linkedin-study-2026/, 673,658 posts; carousels above image and video.
- Buffer State of Social Media Engagement 2026, https://buffer.com/resources/state-of-social-media-engagement-2026/, 52M+ posts; format baselines.

# A carousel people swipe to the end of - Codex task

Expand an approved carousel plan into finished slide text. You have no memory of the conversation that produced it. Everything needed is below.

## You are given

- **{TOPIC}** - the single idea the carousel is about.
- **{AUDIENCE}** - who is swiping, in their own words.
- **{PLATFORM}** - LinkedIn or Instagram. Affects caption length and the closing action only.
- **{PROOF}** - the one real number, story or example the author owns, with its source. The only statistic permitted anywhere in the output.
- **{FRAMEWORK}** - one of: Listicle, Story arc, Before/After, Framework/Steps, Myth-bust.
- **{SLIDE_COUNT}** - an integer from 6 to 10.
- **{VOICE_NOTES}** - how the author writes: their words, sentence length, what they never say.

## Produce

Two files in the working folder.

**1. `carousel-slides.md`**, one block per slide, numbered 1 to {SLIDE_COUNT}:

- Slide 1: hook, **40 characters or fewer**, stating the payoff outright. Then three alternative hooks, each also 40 or fewer.
- Middle slides: headline **50 characters or fewer**, body **25 words or fewer**, then a one-line image note. Exactly one idea per slide. Each ends on an open loop the next slide closes.
- One middle slide, roughly two thirds through, carries {PROOF} with its source named on the slide.
- Final slide: one action only, phrased for {PLATFORM}. On Instagram prefer a comment prompt or a direct question.
- Print the character count in brackets after every hook and headline.

**2. `carousel-caption.md`**: the caption, **200-250 words**, hook inside the first 125 characters, **zero hashtags**, ending on the same single action as the final slide.

## Rules

- Public information only.
- Never invent a fact, a number or a quote. {PROOF} is the only statistic allowed. If a slide seems to need a second number, write `[NEEDS SOURCE]` and move on.
- Anything sent in someone's name says whose name it is.
- No hashtags anywhere.
- Never exceed a count. An idea that will not fit in 25 words is two slides or it is cut.
- Never put two ideas on one slide to save space.
- British English. No em dash characters.
- Match {VOICE_NOTES}. Do not add words the author never uses.

## Return

Print to stdout, in order:

1. The two file paths written.
2. A table: slide number, headline character count, body word count.
3. `HOOK CHARS:` and the slide 1 count.
4. `CAPTION WORDS:` and the count.
5. `HASHTAGS: 0` if there are genuinely none, otherwise list them.
6. Every `[NEEDS SOURCE]` marker left, with its slide number.