Visual design has been the holdout — the creative domain where AI felt most suspect. Designers could point to taste, spatial reasoning, the ineffable craft of making things feel right. But Figma's own team just demonstrated something more interesting than 'AI does design now.' They've shown how AI collapses the middle phase of design work entirely. Not the strategic thinking at the front (what should this do?) or the craft refinement at the end (does this feel right?), but the rushed execution phase where designers frantically translate concepts into pixels under deadline pressure. That's the part Claude Code now handles.
What makes this worth attention: it's Figma practitioners showing their actual workflow, not a vendor pitch. Gui and Alex walk through bidirectional sync between design files and production code via MCP — meaning the Figma file and the shipped product stay coupled, not divergent artifacts. The 90% AI-generated code metric matters less than the structural insight: when your codebase is organized for AI legibility, the tool becomes genuinely useful rather than a novelty that generates plausible-looking garbage.
The frontier isn't 'can AI design' but 'what does design work become when execution speed approaches zero?' Figma's answer: designers move upstream to planning and downstream to craft, spending time on the parts that actually require human judgment. The frantic middle — where most design work currently happens — compresses into minutes instead of days. That's a workflow transformation, not a feature add.
The MCP integration is the real story here — bidirectional sync means I can iterate in code and have design files update automatically, or vice versa. That's not a workflow improvement, that's eliminating an entire class of version control problems that have plagued product teams for a decade.
Show me the designer who's actually shipping 90% AI-generated code to production at scale, not in a demo. The workflow looks elegant until you hit edge cases, accessibility requirements, or the reality that most codebases aren't 'properly structured' for AI legibility. I need to see this deployed across a 50-person product org before I believe the efficiency gains translate.
Notice what they're not saying: how many iterations it took to get Claude to generate usable code, how much prompt engineering expertise is required, or what happens when the AI hallucinates a component that looks right but breaks accessibility. The 90% metric is meaningless without knowing the denominator — 90% of what kind of code, for what kind of features, with how much human correction?
“AI has shifted design work upstream to planning and downstream to craft, eliminating the rushed middle phase of execution”
Why this matters for operators: Product teams rethinking design-engineering handoffs and evaluating whether their codebase structure enables or blocks AI coding tool adoption
I cover AI×GTM intelligence like this every Wednesday.
Get STEEPWORKS WeeklyA good if high level overview
This analysis was produced using the STEEPWORKS system — the same agents, skills, and knowledge architecture available in the GrowthOS package.