Describe any message your product needs to send, from a feature launch to an error to a policy change. Cadence designs it for iOS, Android and web from your own components, then tells you if it's ready or who needs to see it first.
Your rules are the input. So the fastest way to ship is also the on-brand way.
Before the plugin, there was this interactive concept, shown on six fictional Figma launches. Pick one and watch it build and review the whole set. The plugin launching soon is built on the same rules.
⁘ tools menu · ▤ the rules · { } what's underneath · Live the company feed
Three steps, instead of a dozen requests.
Write the message in plain words. Cadence works out what kind it is: an error, an announcement, a milestone, a policy change, or one of seven more.
It picks where the message goes and how loud to be. Push only when it can't wait. Then it builds layered frames from your library: real components, variables and named layers.
A review agent checks every frame against your rules. Ready to publish, needs review by a named team, or fix needed. AI suggests, people decide.
Cadence doesn't own the canvas. It runs as a plugin inside Figma, next to Figma's own AI.
That shaped every decision: be clear about what it made, only touch what it's been given, and leave the file cleaner than it found it. A good third-party agent makes the platform feel stronger, not more crowded.
Clear attribution. Everything Cadence makes lands in one section named “Cadence · [your message]”. Run it again and it replaces that section instead of adding another.
Scoped access. It reads only the libraries and assets you point it at, and talks only to the AI model you choose.
Visible handoffs. When a message needs a person, Cadence names the team and the reason: “Needs review · Legal. Mentions pricing.”
Rules a design-system team already owns. Written once, read on every message.
From no asset at all to the heaviest one a surface can hold.
Eleven comms types. Top wins when two messages collide.
The louder the surface, the more it has to earn.
Same launch. Each person hears what matters to them.
Impression-capped, so nobody sees the same message twice.
Published components, every check passed, nothing downstream has to change.
New pattern, or a change other teams absorb. Not wrong. Just not proven.
Failed a spec check. It would ship broken.
Cadence recommends and says why. You decide where it goes.
The concept is built on figma/sds, real tokens, not approximations. The rules were already published in a format a machine can read. They just needed wiring up. The plugin works the same way with any published Figma library, and writes its rules to a DESIGN.md file any AI agent can read.
One message isn't one design. It's a dozen — across a dozen surfaces, on a deadline, all supposed to feel like one thing.
Today that's a scramble. A banner here, a push there, a dialog someone remembers late. Each built by a different person. Each drifting a little.
The hard part of a design system was never writing it. It's getting teams to adopt it.
Adoption doesn't fail because teams disagree. It fails because following the system is slower than not following it. So make the rules an input, not a document. A team gets a complete, on-brand set in one pass — and the only way to get it that fast is to be on-system.
That matters most for the products moving fastest. They don't need a brand review. They need a sandbox with the rules already loaded, so they can build an idea and ship it the same day. Speed to execution, concept to code.
Comms is the pilot. The mechanism is the product.
Take the rules a company already has, make them machine-readable, let an agent resolve them into finished work. That generalises to every place a brand has to hold across many teams — which is what every enterprise design-system team is trying to solve.
Agents from different makers will share one product.
Figma, Jira, Slack. Who built it, what it can see, and when it hands off should never be a mystery.
Extends Julie's product-communications concept — julietorrisi.com/design-ai