An interface review is most useful when it follows a real task. Start with the journey someone needs to complete, such as signing in or finding a skill, and ask the agent to inspect the states along that journey.
Prepare the task
Provide a preview URL or local app, the key user journey, and target screen sizes. Tell the agent which browser tools are available and whether it may edit the interface.
Work through it with your agent
Choose a review method
Find a skill for accessibility or interface evaluation. Inspect its prerequisites: a skill can guide a browser review, but your agent still needs access to the browser and application.
Walk the interaction
Check keyboard order, visible focus, field labels, error messages, and loading feedback. Repeat the core journey at a narrow viewport and with reduced motion enabled.
Fix and revisit
Prioritize defects that block completion or hide information. After an edit, repeat the same journey and capture evidence that the problem is resolved.
Find a starting skill
Audit a website for accessibility and improve its layout
Task searches use OpenAI for semantic matching. Keep secrets and repository contents out of the search text. Review the suggested skill’s source and instructions before using it.
What to ask for
An issue list tied to specific screens and interaction steps, with before-and-after evidence for implemented changes. Include remaining manual checks rather than treating an automated scan as a complete audit.
Review the result
Confirm that keyboard and screen-reader behavior match the visible interface. Check longer text, zoom, empty states, and error states; a clean screenshot alone does not establish usability.
Continue the workflow
Connect Claude Code or Codex, then copy the pinned agent prompt from a public skill page.
Related workflow: Agent skills for API reviews and better tests.