In a world increasingly mediated by AI agents and large-language models, it’s difficult to know where the people start and where the machines end.
Automation without understanding leads to more friction, not less.
Public agencies have obligations to their residents. Those obligations don’t change because someone has invented a faster way to ask.
When AI stands between people and the services they need, people lose the ability to express needs in their own terms and lose access to human judgment in complex situations. I design for that disintermediation: who keeps authority, where a person can step back in, and what the system is allowed to do next.
AI also changes who can prototype and intervene, and that’s the opening. Used well, it brings teams closer to the problem. Used badly, it produces more repair work, and somebody else is left to clean it up. In most organizations nobody really owns the response experience. Somebody should.
Working with agents exposed a content architecture problem. Interface language, narrative, and the rules that decide what happens each carry a different responsibility. A label explaining a cost needs to be accurate. A character, a customer story, or a quoted source can be partial. Changing one should never change the calculation behind another.
The schema separates those concerns. They can share references without sharing ownership: a line of interface text can take its number from the rule that owns it, so changing the rule changes every sentence that names it. Each content type carries its own writing policy, so a button label and a paragraph of story stop inheriting the same rules.
Much of this comes from established website and application practice: structured content, consistent terminology, semantic components and writing guidance tied to a particular use. That is where it becomes an agent experience primitive. The schema makes the structure inspectable. Content policies make the writing expectations available at the point of editing. The goal is to reduce the amount an agent has to infer and the amount of prior conversation it has to remember.
It ships with a validator, an agent skill and an MCP server, so an agent can check its own work through tools.
Content governance, migration and maintenance, and enterprise content strategy, in talks at Confab Central, Confab Higher Ed and the Design & Content Conference.
The 18F publishing program and website redesign, and the content guide and standards for eAPD.
Context engineering, agent experience, and the authority people keep when agents act for them.
prosekeep, plain-language-default, and a content schema with an agent skill and an MCP server.
Product direction, design teams, and delivery.
Blueprints, research, operations, and the policy behind a service.
The practice that carries the work into implementation.
Sometimes you want advice from someone who has been there before. Have a quick question, an idea to bounce, or want to figure out ways we can work together based on what you’ve read here or a referral? Drop me a note.
Get in touch