Design Outcomes Weekly reflections on the real-time work of supporting a design org.
A Japanese katagami dye stencil: a sheet of dark brown paper cut into a repeating field of hexagons, triangles and bands of brick and wave patterns, every cut-away shape showing white.

A stencil is a ruleset for absent hands that anyone, or anything, can apply.

Why this matters

An AI playbook written before all the answers exist still works, as long as every page describes what’s true today. Here’s how ours is structured, and why we wrote it for our agents as much as for our designers.

Unknown Japanese artisan, Stencil with geometric shapes, 1800–1909. Rijksmuseum, Amsterdam

Write the Playbook for the Agents Too What’s inside our AI UX playbook, why its principles describe what’s true today in place of what we aspire to, and how we’re sharing a work in progress without opening a feedback free-for-all. An AI playbook written before all the answers exist still works, as long as every page describes what’s true today. Here’s how ours is structured, and why we wrote it for our agents as much as for our designers. Write your AI patterns down before they’re final: principles that describe what’s true today, patterns harvested from your most reusable surface, components only once they exist in code. Then post it with targeted asks and a deadline, not an open comment period.

Write the Playbook for the Agents Too

What’s inside our AI UX playbook, why its principles describe what’s true today in place of what we aspire to, and how we’re sharing a work in progress without opening a feedback free-for-all.

For a few weeks now, one of our most senior designers has been writing the playbook for how AI should look and behave across our product: the principles, the patterns and the components we deliver AI capabilities with. This week they walked our design leadership group through it, and the next morning it posts to our broadest product channel, on purpose, before it’s finished.

The anatomy is simple. It opens with the 3 jobs we want AI doing for the people we serve: understand, act, assure. Then come 5 principles, and then the patterns, which are the heart of it: when the AI icon earns a place on a feature, what an AI empty state says, the prompting component and the disclaimers it carries, how suggested prompts appear, how a user’s message renders, what the product shows while the AI is thinking, how a response looks, how sources get cited, and the 2 kinds of rating: a thumbs on each response that improves outputs, and a feature-level ask that exists for a different reason and was deliberately designed to work for any feature, AI or not. The patterns were harvested almost entirely from our most reusable surface, the conversational assistant, because our more bespoke AI features each work one way: a pattern that generalizes has to come from a surface that generalizes.

My favorite edit in the document is a rename. One of the principles arrived as unified and shipped as consistent, because we don’t have a unified AI brand yet, and a principle has to describe what’s true today in place of what we hope gets built. The components section follows the same discipline: it holds only what exists in design and in code, so the shortest section in the playbook is also the one you can trust completely. The author applied the same restraint to their own authority: when the group debated who decides whether a feature gets labeled as AI, their answer was that design shouldn’t claim that call alone, because it’s a product and trust decision as much as a visual one.

The sharing plan got as much design as the patterns did. The post calls out specifically where we want reactions, the principles most of all, and it carries a deadline, because we don’t want to leave a document like this open for feedback forever. It’s framed as a living thing from the first line.

“We’re going to iterate, and we want to make it accessible to everyone, including our agents.”

That last clause is the part I’d steal. The playbook lives in our team wiki in a structure that can hook into the design system, which means the most-used patterns get componentized first, and the agents we’re starting to put into our workflows can read the same rules our designers do. A pattern an agent can read is a pattern that gets applied even when no designer is in the room.

One question stays: when does AI working in the background earn the label? One of our design leaders worried that the AI icons are starting to show up everywhere, and another designer pushed from the other side: anything using AI needs some indication, and we don’t need to be heavy-handed about marking every feature. We didn’t resolve it with a rule, and we’re resolving it with a ritual instead: watch for the cases in crit, add each one to the playbook, and let the rule write itself from examples. For a document whose whole job is describing what’s true, that’s the right way to grow.

Filed under ML Maker's Log — Process, visible.