An AI–Figma playbook teams actually use.

The problem

After three generative-AI features shipped in three months, every new gen-AI team was solving the same component-governance problem from scratch. Variant taxonomy, prompt-to-component flow, disclosure patterns, accessibility gates — all reinvented. Drift appeared by month four.

The problem wasn’t talent. It was that the system needed a shared playbook, and no one had written one.

The process

I spent about six weeks writing and testing the playbook with two early-adopter squads before going org-wide. The content was six sections:

• Component taxonomy. Every gen-AI surface in the product maps to a small, named set of components. New variants live in the taxonomy, not outside it. The taxonomy is the only source of truth.
• Variant strategy in Figma. Variants live in Figma the way they live in code: every state has a reason to exist; every reason is documented in the variant description.
• JSON→tokens pipeline. Generated UI pulls tokens from the design-system JSON, not from Figma frames. The pipeline is one-way; if a designer wants to change a token, they change the JSON, not the Figma file.
• Storybook stories for generated components. Every gen-AI component ships with a Storybook story that covers the variant, the failure state, and the loading state. Stories are reviewed in PRs.
• Chromatic visual review. Gen-AI output goes through Chromatic before merge. Visual diff catches the things code review misses — off-tone color, off-rhythm spacing, off-pace transitions.
• Accessibility and disclosure gates. The same gates that govern non-AI features, applied uniformly. None of this is an AI exemption.

The playbook shipped as a Notion document with embedded Figma, plus a quarterly demo where I walked through one squad’s application of it.

The outcome

Adopted org-wide inside two months of going wide. The interesting signal wasn’t adoption — it was what teams stopped doing. Component drift flattened. The “is this accessible?” debate stopped recurring in new-team kickoffs because the answer was in the playbook. Time-to-first-component for a new gen-AI squad dropped materially. Variance in visual quality narrowed, which made leadership trust the work enough to fund more of it.

The thing I keep coming back to: a playbook nobody uses is decoration. A playbook that stops recurring arguments is compounding. The bar for a playbook isn’t “is it complete” — it’s “does it stop a specific recurring debate?”

This one did. Twice.