Instruments.
Most writing about AI for designers stops at the prompt. Write a better prompt, get a better answer. That's where most of the writing stops.
But the prompt is just the part you can see. There's a lot more going on around it that decides whether the tool is reliable enough to put in front of a client. A schema the agent reads against. A function that runs after the agent and cleans up its output. A rule that's enforced in the build, not written in a doc. A manifest that holds the stuff none of the agents own on their own. A way to turn “a hose clip but for a 25mm pipe” into something the model can actually do something with.
I started calling these instruments. They're not prompts. They're the things around the prompt that make the prompt work. Once you start building them, you stop being a person who uses an LLM, and you start being a person who designs with one.
A few patterns keep showing up. A substrate is what the model reads to get grounded — design system values, geometry primitives, a scene rulebook. The model has something to stand on. A contract is what the output has to do to count as the thing it claims to be. If the agent says it produced a UX audit, the contract is what makes it an audit instead of confident-sounding text. A harness is what the agent runs inside. It bounds what the agent is allowed to emit. The sanitizer that clamps the parameters. The manifest the renderer has to read from. Those are harnesses.
I keep reaching for those three. There are others — memory, provenance, telemetry — and most of the projects below use several at once. The vocabulary helps name what's going on. The list isn't fixed.
Five instruments. Three months of delivery work. The prompt teaches the model what to think about. Everything around the prompt decides whether the work is good enough to ship.
Three projects, three months
Three projects, three tools, three months. The arc that produced everything else in the series. Start here.
I was typing the eightieth chat message into Figma
An at-work agent build for a global consumer brand, two weeks under deadline, and what it taught me about the instrument designing-for-agents actually requires.
The four moves that turn an LLM into a UX auditor
Three versions of a UX audit agent later, the four moves that separate an audit from a list of confident-sounding findings.
Why the UX audit tool had to pass its own audit
Three architectural moves that made the difference between a tool that needed me and a tool that did not.
When the right answer is for the agent to do nothing
A photo pipeline, sixty lines of Python, and the case for non-action as a designed outcome. The sanitizer that catches what the prompt does not.
I don't speak geometry
A 3D printer, a folder of almost-working models, and what one good prompt and three failed ones taught me about the substrate problem.