Claude Carousel Automation

🎠 What this is: a repeatable automation that turns Claude Code + Paper into a carousel factory. You describe the post, Claude researches it, writes a slide-by-slide script for your approval, designs every slide on a real design canvas, scrapes real web assets, and exports finished Instagram carousels (and even animated video slides). This guide covers the one-time setup, then exactly how to run it.

⚡ The 30-second version

1. Set up once — install Claude Code, drop in the framework file + /carousel skill, and connect 4 tools (Paper, Firecrawl, a components library, and a tiny video-export script).
2. Run it — type /carousel. Claude asks a short intake (topic, audience, format, and your 3 reference carousels), researches, and writes a script.
3. Approve the script — nothing gets designed until you say go.
4. Claude builds + exports — finished PNGs (and MP4s for animated slides) land in a dated folder, with a caption.

🧰 Part 1 · What you'll need

🔧 Part 2 · Setup (one time, ~30 min)

Step 1 — Install Claude Code

Install Claude Code (CLI, desktop, or IDE extension) and sign in. Create a dedicated project folder for your carousels, e.g. CarouselWorkflow/, and open it in Claude Code.

mkdir CarouselWorkflow && cd CarouselWorkflow

Step 2 — Add the framework file + the /carousel skill

Two files make the automation repeatable:

• CAROUSEL-FRAMEWORK.md at the project root — the "source of truth" that defines the pipeline, the craft rules, slide archetypes, and your style system. Claude reads this before every build.
• A /carousel skill (in .claude/skills/carousel/) — a short instruction file that tells Claude to follow the framework: run intake → research → get script approval → build in Paper → QA → export.

💡 The skill enforces two hard gates: never skip intake, and never build before you approve the script. That's what keeps results on-brand.

Step 3 — Connect Paper (the design canvas)

Install the Paper desktop app and connect its MCP plugin to Claude Code (the paper-desktop plugin). This lets Claude read and write every node on the canvas — text, color, layout, images. Once connected, Claude can create artboards (1080×1350), lay out slides in HTML, screenshot its own work, and export PNGs.

🔑 The key distinction: Claude can't touch a design file on its own — the Paper MCP is the bridge that gives it hands on the canvas.

Step 4 — Connect Firecrawl (real assets)

Get a Firecrawl API key and add the Firecrawl MCP server to Claude Code:

claude mcp add firecrawl -e FIRECRAWL_API_KEY=fc-your-key-here -- npx -y firecrawl-mcp

Now Claude can scrape a real site and pull back its logo, a full-page screenshot, and clean markdown — so your slides use real web assets instead of placeholder gray boxes. (Verify it works by asking Claude to scrape one site before you rely on it.)

Step 5 — Connect a component / shader library

Wire in a UI component source such as 21st.dev (via the Magic MCP) plus a shader/background library (e.g. shadcn/ui for primitives, ShaderGradient for animated gradients). This gives Claude real, polished components to embed — buttons, cards, gradients — instead of hand-rolling everything.

Step 6 — Set up the MP4 export pipeline (for animated slides)

Animated slides are built as self-contained HTML files. To turn them into crisp MP4s without screen recording, use a tiny headless pipeline:

# one-time, inside your project
npm init -y && npm i puppeteer-core@23
# (requires Google Chrome + ffmpeg installed)

A small capture.js script renders each animation frame-by-frame in headless Chrome at 2× and stitches them with ffmpeg into an exact 1080×1350 h.264 MP4. One command per animated slide:

node capture.js "file://$PWD/slide1.html" slide1.mp4 poster.png 12 30 4
# args: <url> <out.mp4> <poster.png> <durationSec> <fps> <cssOffsetSec>

✅ This replaces screen-recording entirely — it's sharper, exactly the right size, and repeatable.

🚀 Part 3 · How to run a carousel

Type /carousel in Claude Code. The pipeline always runs in this order:

INTAKE → RESEARCH → SCRIPT (you approve) → BUILD → QA → EXPORT

3a · Intake — the questions Claude asks first

Claude batches a few quick questions before anything is designed:

1. Topic + angle — what's the story, and what's your take? (news drop / tutorial / listicle / hot take / case study)
2. Audience + payoff — who's it for, and what's the one thing they walk away with?
3. Hook feeling — should slide 1 trigger curiosity, FOMO, controversy, or "save this" utility?
4. Format — platform + ratio (default Instagram 1080×1350), slide count, and whether any slide is animated.
5. CTA — follow / save / comment a keyword / link in bio?
6. Real assets — any URLs to scrape, screenshots to embed, stats, or handles to credit.

3b · 💬 The reference-carousels question (this defines your look)

This is the most important style question. Claude will ask:

💬 "Can you give me maybe three reference carousels to work off of so that I can get the desired carousel that you want for theme designs?"

Drop in 3 carousels whose style you want to emulate (screenshots or links). Claude studies them and locks a reusable style preset from them — palette, type scale, background/texture, a recurring character or masthead, and the small "furniture" that repeats on every slide. Everything after this stays consistent with those references.

📌 Tip: pick references that actually share a vibe. Three wildly different looks give a muddy result; three cohesive ones give a crisp, on-brand system.

3c · Research → Script → Approval

• Research: Claude pulls live sources (and scrapes real assets via Firecrawl), saving a topic summary with stats + sources.
• Script: Claude writes a per-slide script — for each slide: archetype, headline, support copy, the artifact + its data, and the source.
• Approval gate: you review and approve the script before Claude opens Paper. Tweak freely here; it's cheap to change words, expensive to rebuild slides.

3d · Build → QA → Export

• Build: Claude creates the artboards, builds a consistent masthead/footer once, then lays out each slide, screenshotting and self-critiquing as it goes.
• QA: full-deck pass — overlaps, contrast, alignment, consistent margins, an artifact on every slide, the squint test on slide 1.
• Export: finished slide_1.png … slide_N.png (plus .mp4 for any animated slide) land in a dated folder with a ready-to-paste caption.txt.

🎨 Part 4 · What kind of carousels do you want to make?

Before (or during) intake, decide the shape of the post. These choices, plus your 3 references, fully define the output.

Style

• Visual theme — defined by your 3 reference carousels (Part 3b). Claude derives the palette, fonts, texture/background, recurring character, and repeating furniture from them.
• Density — airy (one idea per slide, big margins) or dense (stacked artifacts, busy margins).
• Voice — cozy/lowercase, or confident/editorial. (Usually follows your references.)

Format

• Slide count — news drops ≈ 7; tutorials ≈ 8–9; listicles = hook + N items + recap + CTA.
• Motion — keep it all static PNGs, or make 1–2 slides animated MP4s (great for launches / hype posts).

Common carousel types

✋ Part 5 · The craft rules (why it doesn't look AI-generated)

These are baked into the framework so every carousel feels authored, not auto-generated:

1. Every slide ships an artifact — a prompt card, UI mockup, real screenshot, chart, or character scene. A typography-only slide is a draft, not a slide.
2. Texture or furniture, always — a background texture or consistent corner/masthead system. Flat default-white reads as generated.
3. A recurring character or masthead — makes the whole feed feel like one author made it.
4. Plausible specificity — real handles, real numbers, real dates. Generic "Lorem / User123" kills it.
5. Imperfection on purpose — slight rotations, an overlapping sticker, a hand-drawn underline, an aside in parentheses.
6. Max 3 typefaces, but extreme scale contrast (tiny mono label next to a giant headline).
7. One accent color does the heavy lifting; dark slabs act as the "second color."
8. Copy sounds like a person — contractions, asides, second person; never "In today's fast-paced world."
9. Banned: purple gradients, Inter-only, centered-everything, symmetric 3-icon grids, generic 🚀🔥, drop shadows on text.
10. The squint test — at thumbnail size, slide 1 must still read its 3–5 word promise.

🧩 Part 6 · Slide archetypes & story arcs

Claude picks one archetype per slide (each has a known layout recipe):

• Cover / Hook — big claim + one visual anchor + a teaser line
• Setup / Problem — short headline + an artifact showing the pain
• Prompt Card — a dark mono card with the actual copy-pasteable prompt
• UI Mockup — a realistic product interface built in HTML
• Stat / Chart — one hero number + a real bar chart
• Comparison / Diagram — stacked cards connected with arrows
• Listicle Item — number + name + image grid + one-liner
• Quote / Proof — a pull-quote slab with real attribution
• Steps / Recipe — numbered rows, one line each
• CTA — the ask (save / follow / comment), character, what's next

Story arcs:

• News drop: Hook → What it is → Why it matters (numbers) → Proof → What you can do → Catch → CTA
• Tutorial: Hook → First step → Build pages (×3–4) → Result → CTA
• Listicle: Hook → Item ×N → Recap grid → CTA

🎬 Part 7 · Animating slides → MP4 (no screen recording)

Instagram carousels can mix video and image slides. The workflow makes the thumb-stopping motion slide painless:

1. Claude builds the animated slide as a self-contained HTML file (same 1080×1350 design, with CSS/JS animation — a logo drawing in, a character bouncing, a gradient drifting, or a real screen-recording playing inside a browser window).
2. The capture.js headless pipeline (Part 2, Step 6) renders it frame-perfect and exports a crisp 1080×1350 MP4 — no manual screen recording, exact dimensions, repeatable.
3. Post the MP4 as that slide, PNGs for the rest.

🎥 You can also drop a real screen recording into a slide (e.g. show the deck being edited live) — Claude composites it into the slide's frame and exports the MP4 the same way.

🛠️ Part 8 · Gotchas & troubleshooting

• Always approve the script first. It's the cheapest place to change direction.
• No slide counters. Don't add page number pill — Instagram already shows this.
• Backgrounds must be full-bleed. If a texture shows a grid/checkerboard at the edges, it's a transparency cutout — scale the background up and center it so the margins crop off-canvas.
• Inline colored words can flatten when laying out text — split accent words into their own text nodes so the color sticks.
• Verify Firecrawl before relying on it — run one scrape to confirm the key + connection work.
• Animated-slide filenames from macOS screen recordings contain a hidden character — copy them with a wildcard/find, not a literal path.

✅ That's the whole system. Set up once (Part 2), then every post is: /carousel → answer intake (give your 3 reference carousels) → approve the script → collect finished slides. Want a tutorial, a news drop, or a listicle next? Just say the topic and hand over your references.