From Flat Images to Editable Canvas: Rebuilding an AI Carousel Workflow with Paper Design

TLDR

The author of AI Maker rebuilt their LinkedIn carousel generation workflow to use Paper Design's MCP server alongside Claude Code, replacing flat HTML exports with fully editable design canvases. The new system lets Claude handle content structure and initial layout while the human polishes the last few details directly on the canvas. This is a genuinely interesting model for AI-assisted creative work: let the AI do the heavy lifting, then hand it off to human taste.

The Problem With "Good Enough" Automation

Here's what gets me about this piece: the author had a working system. Claude could read a newsletter post, break it into carousel slides, generate HTML and CSS, embed a profile photo, and export clean images. That is not a small thing. Most people never get that far.

But they kept noticing something. The output was technically correct and visually consistent, but it had a ceiling. Every slide shared the same rhythm. The headlines sat in the same spot. The spacing never quite breathed the way a real designer would make it breathe. And when something felt off, the only fix was to prompt Claude again and hope the next attempt was closer.

That is the part worth sitting with. Prompting for design corrections is not the same as editing a design. Taste is applied differently. You can write "the headline feels too large" but what you really want to do is grab it and move it two pixels to the left. Those are not the same action.

What Paper Design MCP Actually Does

Paper Design is a design tool that runs a local MCP server through its desktop app. Claude Code connects to that server and can create artboards, write HTML into them, duplicate slides, swap text, and update styles, all inside an editable canvas. The output is not a flat image. It is a set of layers you can click into and adjust.

The workflow becomes: Claude creates the structure, Claude applies the brand rules, Claude builds the first version, and then you touch the design directly for the final 5%. That split is worth more than it sounds.

The author compares this to Figma's MCP, which is fair because Figma is the obvious default answer for anyone who works with design tools. Figma's MCP can create native objects, reference components, and work with existing design systems. For product design that has a library full of reusable components, Figma probably wins.

For editorial content like LinkedIn carousels, Paper fits better right now. It has no custom font support gaps, no 20kb response limits, no full-seat requirements for agent write access. Claude is already good at HTML and CSS. Paper just changes where that HTML lands, from a browser renderer into an editable canvas. That is a genuinely clever architectural choice.

The Honest Trade-offs

I keep thinking about the cost question here. Paper's free plan gives you 100 MCP tool calls per week. That runs out faster than you'd expect if you're generating multi-slide carousels. Paper Pro is $20 a month, or $16 with yearly billing.

Whether that math works depends entirely on how often you create this kind of content. The author makes a point that Paper isn't just for LinkedIn carousels. The same setup works for landing pages, lead magnets, Instagram carousels, visual explainers. If you're regularly producing that kind of output, the tool starts paying for itself in time saved. If you post a carousel every couple of months, it probably doesn't.

There's also a setup cost that's easy to underestimate. You need Paper Desktop installed, the MCP server connected to Claude Code, a brand configuration saved, your profile photo accessible. The author recommends testing the connection by asking Claude to create a red rectangle before touching anything serious. That is good advice. It is the kind of small debugging step that saves you 30 minutes later.

Why This Workflow Pattern Matters

The piece ends with something I think is the actual insight: "Not another AI image generator. Not another rigid template. A system where Claude gets me to a strong first draft, Paper keeps everything editable, and I can use my own taste to finish the last few details."

That framing is useful beyond LinkedIn carousels. The best AI-assisted workflows are not the ones where the AI produces a finished artifact. They're the ones where the AI gets you to a strong starting point fast, and then you take over with judgment the AI doesn't have. Human taste applied at the end is not a workaround. It is the design.

The addition of Typefully's scheduling integration completes the picture, from idea to editable carousel to scheduled post, without leaving Claude Code. That is a real workflow, not a demo.

I Rebuilt My LinkedIn Carousel System, It Now Looks Like A Designer Made It

Key Takeaways

• AI-generated design output works best when it produces editable artifacts, not flat images, so human taste can finish the job
• Paper Design's MCP server lets Claude Code write directly to a design canvas with real layers, not just render HTML to a screenshot
• The workflow split is: Claude handles content structure and brand rules, human handles the final polish
• Figma's MCP is stronger for component-library-driven product design; Paper fits editorial content creation better right now
• Free tier (100 MCP calls/week) is enough to test the workflow, but regular use likely requires the $20/month Pro plan
• The pattern generalizes beyond carousels: any creative output benefits from AI doing the heavy lifting and human taste doing the finishing