Claude sucks at carousels. This is the system that fixes it.

Plan every slide

Give it an idea, get back a complete slide plan: slide count, exact text per slide, and a one-line visual direction for each. It enforces the rules I learned the hard way:

• Hook = the pain point, zero setup. Two lines max, no subtitle explaining what's coming.
• No framework slide. Slide 2 delivers the first payoff — it never restates the promise.
• One idea per slide. If you need “and” to describe it, it's two slides.
• 25 words max per slide. Long text is what image models mangle — anything longer becomes a diagram.
• Every slide specifies its visual. Rules and comparisons live in diagrams, not paragraphs.
• The CTA gets hook treatment. One action, stated plainly.

Give Claude taste

Image models have no design opinion — they'll happily give you cramped text, three accent colors, and gradients from 2021. So I hardcoded my taste into a second skill that injects a design system into every image prompt automatically: generous margins, one dominant element per slide, one accent color used sparingly, bold condensed titles, visual metaphors over paragraphs, and recurring furniture (a step label, a page arrow) on every slide.

Connect Claude to an image model

People overcomplicate this. Claude does the thinking (copy, prompts); the image model does the rendering. The bridge can be as dumb as two browser tabs. Three options, easiest first:

• A — ChatGPT (start here). GPT Image 2 is best-in-class at text in images and it's right inside ChatGPT, free tier included. Claude gives the prompt, you paste it, attach references, generate. No API keys.
• B — Higgsfield MCP (what I use). An MCP connector lets Claude call the image model directly — prompt, generate, iterate, all in one chat. A bit more setup, multiple models in one place.
• C — API routes (for builders). In Claude Code, wire gpt-image-2 via the OpenAI API or fal.ai / Replicate and script the whole pipeline. Overkill for your first one; nice once you're producing weekly.

Generate 3 hooks, not 1

The hook is the most important image in the carousel, so it never gets one attempt. Collect 2–3 reference slides you love (to steal the feeling, not copy), have Claude write 3 structurally different hook prompts — different layouts and metaphors, not color swaps — generate all three with your references attached, then iterate on the one that stops your own scroll.

The 3 hook variants for this exact carousel — same copy, three different visual bets. Pick the one that stops your own scroll, then iterate rounds on that one.

• Misspelled text? Cut words or go bigger — don't just re-roll.
• Busy layout? Remove one element. Diagrams need three components max.
• Two to four rounds is normal. On round seven? The prompt is wrong, not the model.

Lock the anchor

This is what makes a carousel look like one designer made it, not six random generations stapled together. Run the same 3-variant process on your first value slide, pick the winner — that image is now your anchor. For every remaining value slide, attach the anchor as a reference and prompt: match this layout, typography, palette, and spacing exactly; change only the text and the diagram. If a slide drifts, re-attach the anchor and add "do not change the layout, type, palette, or margins."

The CTA is a hook

The last slide is the one people share, screenshot, and act on — so it gets the exact same treatment as the hook: 3 structurally different variants, iterate, pick the winner. Keep it in the same visual family, but let it break the layout for emphasis. One rule on the copy: one action only. Comment a keyword, or tap the link, or follow — never two. Every extra option cuts your conversion.