How a VGV Designer Builds Custom Figma Plugins with AI

Designer Murphy O'Rourke leads design on some of Very Good Ventures' most complex programs. He describes himself as a more technical designer than most, and the background backs it up.

He started in front-end work, learned HTML, CSS, and JavaScript, taught himself enough WordPress to build sites for friends and clients, and picked up Git along the way.

Recently, on the Very Good Engineering podcast, he shared how he uses AI to build an entire ecosystem of custom Figma plugins, from his first ChatGPT experiment to Claude Code.

Plugins vs. the Figma API

A Figma plugin is a way to interact with the canvas and the data behind it. Murphy draws a line between two things people often confuse.

The Figma API is more powerful but sits behind higher subscription tiers, so reaching features like variables requires an enterprise or org plan. The Figma plugin API runs locally in the app, uses your own authentication, and opens up a huge amount of that same functionality.

The payoff is ownership. Instead of stitching together marketplace plugins that each get you ninety percent of the way, a team can build tools shaped exactly around how it wants to work.

As Murphy puts it, you no longer have to keep up with the Joneses and work the way everyone else does.

From Flutter Styles to Control R

Murphy's first plugin, Very Good Flutter Styles, shipped in late 2023. It takes the theme colors and text styles from a Figma component library and exports them as Dart code.

That solves a real handoff problem. A project can carry hundreds of styles, with each text style holding six to ten properties, and moving all of it into code by hand is slow and error-prone.

He built it by talking to ChatGPT, asking for a function to pull the style data, pasting the result into a JavaScript file, and wiring the pieces together with his own limited knowledge of code.

It took about a month of working in fits and starts, plus close collaboration with VGV developers to get the output up to standard. The result drops clean Dart into the codebase, with each style commented and organized.

Most of what Murphy builds traces back to VGV's template files and how they handle variables. The team builds adaptive layouts from variables and uses Figma modes to flex color and size, so a single component can render as a phone, tablet, laptop, or desktop, in light or dark, at different text and global scaling levels.

One of his newer tools started as a Claude skill that interviewed the user and customized a template file. It worked, and it ate an entire session's worth of tokens. So he rebuilt the same idea as a plugin: the Very Good Design Kit.

• It validates that a file's collections are structured correctly
• It walks through brand colors and multi-brand support
• It configures font families, type scale and ratio, and numbering systems
• It generates a complete, project-ready Figma template

The plugin he is proudest of is Control R, for renaming variables. It gave designers search-and-scope renaming for variables before Figma shipped its own version, borrowing the pattern Figma already used for renaming layers.

It picked up traction on design blogs he had never heard of and shout-outs from people he admires in the design community.

AI Changed the Workflow

Murphy's toolkit has changed a lot since 2023. He moved from ChatGPT to Copilot and Cursor, and now leans on Cursor and Claude Code. A proof of concept that once took a month can now take fifteen minutes to an hour.

He keeps a public starter kit on GitHub with the boilerplate files a plugin needs, common functions, and a set of Claude skills and Cursor rules for chores like generating a README or a comment manifest.

For quick ideas he reaches for Cursor and simply tells it to build. For anything complex he opens Claude Code in plan mode, often planning with Opus and building with Sonnet, clearing context between the two.

He estimates plan mode one-shots the idea about ninety percent of the time, and that it sometimes thinks of things he had not considered. He also points the agent at the Figma file through the Figma MCP server.

The conversation keeps circling back to a theme VGV calls trust and taste. With these tools, the old lines between product, design, and engineering get blurry.

A designer can take a first pass at code, an engineer can prototype in a design tool, and the back and forth gets faster. Murphy is clear-eyed about the limits. He can jump into code, but he leans on developers to judge whether it belongs in production.

He points to shaders as an example. Custom shaders and interactive effects used to demand a lot of specialized knowledge and time. With AI explaining them in plain language, they open up to more of the team, which makes room for work that used to get cut for budget or schedule.

Murphy O'Rourke's plugins show what happens when a designer treats AI as a way to build the tools he always wished he had. The work is less about replacing engineers and more about closing the gap between design and code, so a team can move faster without giving up structure or accessibility.

For anyone curious about extending Figma, his advice is simple: find a repetitive task, learn just enough to automate it, and let the AI handle the parts you do not know yet.