ChatGPT now has Sketch (type “@Sketch”, draw a rough layout, get it rendered), templates for formats like posters and merch, comments pinned directly onto an image for focused edits, and shareable prompts so someone else can run your idea on their own photos. That’s a design tool’s vocabulary showing up inside a chat. Generation latency is reduced by up to 50% compared with Images 2.0, it’s live on every ChatGPT tier including free, and developers get two new API models: Flare as the default, Sunburst for precision work with longer generation times.

“I worry a little bit about function following form.” That’s Katarina Batina, VP of Design at Shopify, in the middle of a piece about how completely her team reorganized around AI tooling. Roughly 60% of Shopify’s designers now push code directly to production, up 38% since last September, with over 5,500 live prototype sites built on Quick, the internal hosting platform. Her caveat is the part worth keeping: ​“There’s so much discourse around what tools are in fashion this week versus next week. The problem is we risk losing connection with the questions we’re trying to answer and the people we are building software for.”

The skills demo is the one to watch. Rodrigo calls a slash command for a skill his teammate Tammy wrote in plain English, carrying her own design-review conventions, and the agent returns the critique as annotations right on the canvas. Team practice becomes a callable command.

Miggi rounds up ten community skills for the Figma agent, though the real news is in the first line: you can now author skills inside Figma. Ask the agent to draft one using a design frame as reference, preview it, then keep iterating in chat or open the markdown editor for manual changes. Emil Kowalski’s /find-animation-opportunities, which scans a file for places that don’t animate but should, and Jakub Krehel’s /better-interface which catches the details your interface is missing, are the ones I’d install first. 

The post above shows the skills, this one shows the plumbing. Miggi builds 3 skills: a type scale generator that also writes the variables, an image audit that drops a DPI card beside every image, and Shader Starter, which pairs neatly with the shaders update below.

Everything above is agents arriving in Figma, but this one runs the other way. Nate Parrott, the Anthropic designer behind Claude Design, put its artboard canvas inside Claude Code, so options get drawn where the code already lives. One brief, several editable artboards, and then Claude implements the one you pick. /design-sync can load your design system from a repo or your local codebase first. 

Jake Albaugh, Developer Advocate at Figma, points a fleet of agents at an entire component library, then reuses the same agents as validators: each file re-fetches from Figma and surfaces what drifted after a property rename. His real argument for Code Connect is context economy – without it the MCP server hands the agent every style value it might need, and compaction is lossy.

Detach the agent chat into its own window on the desktop app, move it anywhere, keep it in view while you design. Figma expects you to be watching an agent work while you do something else, which turns the chat from a sidebar into a persistent companion window.

“MCP gives your AI reach. Skills give it a recipe.” Murphy Trueman compresses two acronyms into one sentence you can actually hold onto. MCP is access: Claude reading your real Figma file, walking every fill on a button set and finding the fourteen layers still on raw hex, all of them on hover and disabled states. A skill is the playbook it follows once it’s in there. Their rule for writing one is the third time you catch yourself explaining the same preference, write it down.

Laura Fehre, Designer Advocate at Figma, demos three plugins she generated for herself: one that updates a component description and drops a changelog onto the canvas after a variant change, a print preflight that checks for 300 DPI and maps RGB to CMYK with die and safety lines, and a translation pass that duplicates screens into German, Dutch, and Japanese to catch overflow before shipping. The print one is the standout, since Figma has never handled prepress natively.

Admins can now set a custom AI credit limit for an individual person, and when someone hits that limit they can send a request to their admin to raise it. Design tools now ship with an expense approval flow.

Kaelig Deloumeau-Prigent, co-chair of the W3C Design Tokens Community Group, audited 20 open-source design systems and 6 platforms, logging 187 AI affordances and 157 techniques. 17 ship an official MCP server, 17 ship official agent skills, 14 publish llms.txt. Two ship Figma Code Connect: Carbon and Primer. Design systems are being rebuilt as machine interfaces right now, and the design side of that bridge is the part almost nobody has wired up. Code Connect is still Organization and Enterprise only.

“And suddenly, you’re just feeding the AI prompts. And eventually you’ve lost control. And you surrender to the AI.” Dylan Field on TBPN, right after Figma’s Q2 earnings, describing what CPO Yuhki Yamashita named quiet surrender. You start with a picture in your head, the model talks you into its direction, and you never notice the swap.

Figma’s own three-minute primer on getting good output from the design agent. Five habits: write specific prompts instead of ​“make this better,” select the exact layer or frame to scope the edit, build with auto layout so the agent can read your structure, feed it full production-ready UI examples rather than lone base components, and keep a separate chat thread per task. The one most people miss is context hygiene — the agent reads your whole file, hidden layers, unused elements, and old versions included, so clean it up before you prompt.

Where the video above is Figma’s polished pitch, Mal shows the messy middle. She builds her own color-audit skill for the Figma agent, feeding it brand guidelines and example palettes so it flags off-brand elements by severity and annotates the exact violations right on the canvas, then has it rebuild the frame in every approved scheme. ​“Truly a diva just sharing her work in progress,” and better for it.

Admins can now download a CSV showing credit use in beta AI features, giving a fuller picture of AI spend and something to forecast against. Organization and Enterprise plans only.

A large survey of tech workers in 2026 finds the industry splitting into two groups: roughly half feel amplified by AI, while 14% feel destabilized and 12% are simply resentful. Designers and researchers are overrepresented in the fragile half. What’s striking is that the fear isn’t ​“AI will take my job” — only 22% name that. The bigger worry is unsustainable pace and doing more work for the same pay. Career optimism is down, burnout is up, and 53% would actively discourage someone from entering their field.

Dylan Field responds to Gal Shir’s ​“quitting design” post that made a splash in the design community this week and supports findings in Lenny’s survey. He points to the recurring psychological loop that plays out every time a new AI model drops: existential crisis, experimentation, recalibration, repeat. His standing argument is that the attention economy makes design more valuable, not less, because anyone can prompt their way to average, and average doesn’t stand out. ​“This is the moment to be more bold, to take more creative risk, to double down on the power of design.”

Figma’s third annual AI survey, covering 8,403 product builders across 10 markets, lands on a deceptively simple conclusion: AI is most valuable when it’s a team sport. Two years ago, 7% said AI meaningfully changed how their teams collaborate – this year, that number is 41%. The most cited reason is the canvas — where teams can actually riff together rather than trading prompts solo. The cross-functional blurring numbers are striking too: designers participating in development jumped from 21% to 41%, developers doing design work from 44% to 60%. The role boundaries are dissolving faster than most teams have figured out what to do about it.

Contra Labs put four frontier AI models through a rigorous landing page benchmark: nine professional designers, 40 live HTML artifacts, 540 pairwise matchups across typography, layout, palette, and grid. Sol (GPT 5.6) dominated loosely-specified briefs with an 81–83% win rate, praised for capturing tone the brief ​“only gestured at.” Fable (Claude 5) flipped the result on structured briefs, jumping from 31% to 72% client-readiness when given detailed specs. The real finding is that these models have different philosophies: Sol fills ambiguity with its own taste, Fable waits for yours.