The sandbox repo I argued for in Designers Need a Sandbox Repo isn’t only an infrastructure argument. It’s a learning argument.
When designers work in a 1:1 clone of the production repo, using real components, real lint rules, real token values, the feedback they get from an agent is grounded in reality. The agent isn’t working in a vacuum. It’s working in a system that approximates the real one. When something goes wrong, it goes wrong in a way that teaches you something about real constraints, not about a fake approximation of them.
Most people designing with AI agents are doing it in conditions that don’t reflect how they’d actually use them. Prompting into the void, on tools that churn out default themes and interchangeable styles. Building things with no constraints, no users, no stakes. And then they’re surprised when their first real project doesn’t go the way the demo did.
Build with real constraints from the start. Get an understanding of the front end, even if it’s served by a provisional backend. The friction you hit on the way, the errors you catch in the UI, the usability and accessibility you find missing — those are the skills building with AI can hone, and they’re what make you better at design.
Also in August: Build to Learn AI · What Building Teaches That Reading Cannot · The Skill Under the Skill
Daniel Arevalo is a Product Architect and strategic product design leader. He helps product teams turn AI capability into trusted, adoptable workflows, with a background in industrial AI, IT/OT cybersecurity, and high-stakes enterprise software for critical infrastructure.