Teach the agent your system once
The useful AI design work this week is not another model. It is a place to put your spacing, components, and CMS rules so the agent stops asking you to restate them.
Most of the AI noise this week was about models getting cheaper or faster. Fine. What I kept from the shelf was quieter: tools that let you write the rules of your product once, and have the agent reuse them on the next task.
Framer shipped Skills. You save design-system instructions, writing style, and how the CMS is set up. The agent points at your real pages, components, and styles instead of inventing a new spacing scale every time you ask for a section. shadcn/lint went the same direction for Tailwind systems: you say what is allowed, and when an agent breaks a rule the error tells it how to fix the break with your components and theme. Linear put its notification inbox on MCP, so an agent in the editor can clear the queue you already trust instead of opening a second tab and guessing.
None of that is a new look. It is a handoff.
What changed
For a long time the cost of using an agent on real product work was re-explanation. You pasted the token names. You named the components that already exist. You described the CMS shape again. The agent produced something plausible and slightly wrong, and you spent the saved time putting the system back.
Skills, lint rules, and MCP inboxes cut that loop. The product knowledge lives next to the agent. You still judge the result. You just stop typing the same paragraph.
A three-line test
Before you prompt, write three lines only you can write:
- What already exists that the agent must use.
- What it must not invent.
- How you will know the screen is finished.
If you cannot fill those lines, the agent will invent the answers. That is the old prototype tax under a new name.
I am collecting the week's signal for designers who already work this way at shader.md. Five minutes on Friday. The few things that actually mattered.
