Figma turns into a wasteland for tiny AI-powered crews

Running a two-person outfit with a designer cofounder, I've hit a ceiling where our process now drags us down because AI moved the choke point from building to keeping things consistent. The more we lean on AI-generated frontend code, the more minor UI bugs and visual mismatches pile up. These are trivial fixes an LLM agent could knock out instantly, yet most agentic coding tools target developers. My designer can't just fire up a heavy dev environment to tell an agent to nudge padding by 4px or correct a hover state. So we spin in a loop: designer spots a bug, tells me, I prompt the AI, and everyone's time evaporates.

Worse, our source of truth has vanished. We began with a clean Figma library, but now that we ship at speed, I usually push changes straight into code. Once the designer green-lights the live version, we release. Figma never catches up to production. The design files have become dead artifacts, and our actual production component system is fracturing into a mess because we iterate in the browser instead of the design tool.

For others running lean teams, what does your AI workflow look like? I'm hunting for a real-world AI workflow that doesn't force the developer to play human proxy between designer and codebase. If you've gone code-first, how do you stop the UI from turning into a sprawling disaster of inconsistent components? I need a practical tutorial or a set of habits that let designers touch the implementation directly without a CS degree to configure their IDE. Maybe we should double down on a strict design system the AI must obey, but locking that down feels like the opposite of the speed we're chasing. It seems we need a bridge between the vibe-coding phase and a sustainable production codebase.