A test set of prompts that shows how ready your design system is for agents

An agent can answer the same prompt differently each time. One run can produce a clean settings page and the next a gray that is not in your palette. Judging a design system, a rules file or a new model on one screen is judging the dice. A test set fixes the questions, so the only things that change between runs are the ones you changed. Anthropic’s own guide to evaluations makes the same case: success criteria that are specific and measurable, and test cases drawn from the tasks you actually run, edge cases included.

Write prompts the way your team would ask, in 5 kinds:

• Everyday screens. The pages and parts your team builds most often.
• Choices the rules decide. A delete action, a form error, a primary and a secondary action side by side: places where your guidance says which variant to use.
• Gaps. Something the system does not have. The right answer is built from existing parts, or a question back to you, not a new component.
• Edge cases. Long names, empty lists, errors, dark theme, a narrow phone screen.
• Temptations. Prompts that invite a raw value, such as “make the button bright orange”. A good agent says no, or finds the design token that comes closest.

Give every prompt the same short list and the same points, 2 for a pass, 1 for part, 0 for a fail, as the ranking does for systems:

• Design tokens only. A checker finds no color, size or font written by hand. The Simbolik Design System’s checker lists each one, and Stylelint’s color-no-hex rule flags hex colors in any CSS.
• System components only. Nothing rebuilt that the system already has.
• Rules followed. The specific lines of your guidance that the prompt tests.
• Accessible. An automated check such as axe-core passes. Its makers say it finds about 57% of WCAG issues on average, so a person looks at the rest.
• Looks like you. A person answers yes or no. A saved screenshot to compare against, as Playwright keeps, makes that answer quicker and steadier.

Markdown

1 2 3 4 5 6 7 8 9 10 11

## T07: Delete a project

Prompt: "Add a way to delete a project from the project settings page."

Must be true:
- Button in its danger variant, confirmed in an alert dialog.
- The dialog names the project and says the deletion cannot be undone.
- The checker reports 0 values written by hand.
- Focus moves into the dialog, and back to the button when it closes (a person checks).

Score: 2 if every line holds, 1 if only the checker's line holds, 0 otherwise.

Run the whole set when the design tokens change, when the guidance changes, when a component is added, and when the agent or its model gets a new version. Keep the same settings each time and record the totals by date. A drop after a change points at the change. A custom AI-ready system can ship with its own test set; Work with Simbolik (opens in a new tab) if yours should. To put the set to work on a real project, go on to a bounded pilot with an AI agent, and for the checks themselves to acceptance checks. The best AI-ready design systems, ranked is where the systems themselves are compared.