“We’ve increasingly seen that specialized post-training can compete with frontier models.” Figma explains the model behind the agent’s everyday edits. It started with supervised fine-tuning on synthetic edit requests in Figma’s native file format, and an early Llama 3.3 70B trained on more than 250,000 examples was competitive with the top models of the time. The agent now routes the work. Third-party models handle image generation, vectors, and new explorations, while Figma’s own model handles precise edits and iterations.
Amichai Mantinband, co-founder and CEO of Brilliant, read Figma’s model post and lays out where it’s smart and where it might break. Small fine-tuned models are much cheaper and keep getting better, but “Fine tuning teaches taste and figma’s data model, HOWEVER — it does not add intelligence.” Design data is verbose, and real projects have hundreds of components. His conclusion is a trade-off between a frontier model that “can’t speak figma” and a small model that can’t reason past a certain level of complexity.
“I was shocked that 3/4s of designers work in coding agents rather than design tools.” Maayan Albert, formerly at Browserbase, Apple, and Google, is building Proto, “the designer’s IDE”: the shape of Claude Code and Cursor, “but with every design feature you wish they’d ship.” It’s in early access.
What does an AI-ready design system look like in practice? Kaelig Deloumeau-Prigent walks Brad Frost through his updated State of AI in Design Systems report (first covered in #258) and singles out Meta’s Astryx for running nightly evals to check that its skills produce good outcomes. Then they run the audit on Eddie. Brad’s takeaway: the tooling surfaces the gaps, and people still have to decide what matters.
“Designers don’t need help designing, they need help with everything around it.” Jaycee Day’s plugin analyzes your Figma files and writes up what you’ve worked on recently, plus a changelog, for weekly reports, standups, and developers who want to see what changed. She built it after talking with 40+ designers, and in a follow-up post she describes how their workflow shifted: “designers are exploring in Claude, iterating in Figma, then bringing it back to Claude to finish building.”
Click an object in an image and Figma selects it, ready to isolate or erase. It’s a small change to the Select area tool, but it removes the most tedious part of image cleanup. Only on the macOS desktop app.
On the All-In podcast, Dylan Field takes back his own Config prediction: “I said the ceiling will be raised by AI… That was wrong.” AI lowers the floor for making code, motion, and design, but the ceiling only rises when designers “choose to focus, invest their time, and to be bold.” With exponentially more software and flat app usage, that is now where products differentiate.
Figma’s guide to spending fewer AI credits with the agent: scope prompts clearly, attach only the context the task needs, and start a new chat when you switch tasks, since long chats raise the cost of every prompt. For repeated jobs, turn the work into a generative plugin — building it costs credits, but every run after that costs zero.
Enterprise admins can now set default per-user access to paid AI credits at the organization and billing group levels, and manage credit access and limits through SCIM. They can also block people from creating Starter or Professional plans with company emails, and use new admin REST APIs to manage teams, folders, and files. Enterprise plans only, rolling out in the coming days.
“Intelligent UI is a step toward a ChatGPT that can shape an interface around what you’re trying to do. Over time, we want ChatGPT to create more of the interfaces people need for everyday tasks. You should be able to describe what you want to accomplish and work with an experience built around it.” OpenAI’s take on generative UI ships to all ChatGPT users, built on a component library plus a model trained to judge layout and interaction.
Gemini 4 Argon, Google’s latest model, is now in the Figma Make model picker on all plans.
Figma’s newest product is a Markdown editor. “You can think of Specs as a container for your context.” Specs live on the canvas and as their own file type, are multiplayer with comments, and can be filled by hand or through MCP and GitHub. Rolling out in open beta on paid plans.
Main-stage talks from Lenny & Friends Summit, with Marty Cagan on strong opinions loosely held and Claire Vo on whether any of the extra shipping matters. Start with Katie Dill of Stripe, who warns that AI building could leave us with “zombie UI that’s monotonous, vacant, or uncared for.”
Peng Zheng’s personal site updates itself, as he shows host Claire Vo on How I AI. Send his check-in bot a photo, a screenshot or just the name of a place, and it looks the place up through the Google Places API, fixes the perspective, removes people and generates light and dark 3D miniature versions for the site.
Brett McMillin, Designer Advocate at Figma, runs a skill that pulls user personas from a Notion page and critiques selected frames as one persona, in his demo a fintech app for teens. The findings land as annotations on the canvas covering comprehension, findability, visual hierarchy and trust, and the skill is told never to touch the design. “Nothing will take the place of actual user testing and talking to our customers.”
Why stay in the details when agents can do the hard part? Geoffrey Litt, design engineer at Notion, says the time spent in the nooks and crannies of a problem loads your memory with the raw material your brain recombines later, even in the background. His choice for the AI era: get lazy and settle for good-enough, or work harder than ever. “You can already start to tell who is taking each path.”
A checklist for spotting AI-generated interfaces, written after a college app’s “minor UI improvements” update. Gradients everywhere, pulsing “active” badges that can never be inactive, fingernail cards, glassmorphism, and taglines built from “Elevate”, “Seamless” and “Unleash”.
Anton Lovchikov, Head of Design at Evil Martians, makes the case that “a designer’s main job isn’t drawing screens. It’s pulling rules out of chaos and making sure they’re followed.” His proof is a 700-line localization refactor, after which Claude rewrote half the strings. The fix was a three-layer system of product principles, copy rules traced back to them, and skills built on those rules. Now when he dislikes a string, he asks Claude which rule produced it and fixes the rule.
Evil Martians’ eight-part framework for keeping AI-built interfaces consistent. Component contracts say when to use and not use each component, an inventory lets agents shortlist components before reading any code, and a design linter fails the build on a raw hex color. The agent skill is published too, split into craft and use workflows.
Jeff Smith of Fieldwork benchmarked frontier and open-weight models on adding icons to an existing set, with humans grading recognizability, process, consistency and execution. His summary: “they’re steadily improving execution, their choices often overly literal.”