Scope, in three words
The most useful thing you can say to an AI co-pilot is what not to touch.
2026-04-25 · 3 min read
I asked the agent to add motion to a component. It added motion. It also rewrote the component. It changed the color, restructured the layout, renamed three props. The motion was fine. Everything else had to be reverted. The fix was a three-word prompt I had to learn the hard way. Motions only.
AI co-pilots improve by default
They are trained on the implicit goal of being helpful. Helpful, to the model, often means more. More refactoring. More cleanup. More edits adjacent to the one you asked for. On a junior designer this would be ambition. On an agent it is a scope problem. The default direction is up. Yours has to be sideways.
Specify the negative space
Telling the agent what to do is half the prompt. Telling it what not to touch is the other half. Do the motions, not the design. Fix the bug, not the file. Update the copy, not the layout. Three extra words at the end of every prompt save twenty minutes of revert at the end of every task. Boring discipline. The most expensive habit to skip.
Why this is a design problem, not a prompting problem
A design tool that improves what you did not ask it to improve is making a contract decision on your behalf. The contract should be visible. A plan view, a confirm step, an explicit boundary, anything that shows the agent what is in scope and what is out before edits land. Until that exists in the product, three words at the end of every prompt does the job.
Reach out: jason@theft.studio