Why Prototyping Still Matters When Building With AI Agents

Summary

Emm’s author argues that prototyping remains necessary when AI agents are used to design complex, multi-step product workflows. The example is Emm, a local-first notes manager and issue tracker for agents and humans, where projects can exist locally before cloud backup and syncing also involves signed-in users, account changes, invitations, and a highly important first-time experience. In the first attempt, the author asked an agent named Astra to gather requirements, propose options, write a specification, and implement it. The agent handled much of the underlying plumbing, but produced a cumbersome flow that required about seven clicks where four would have been enough and sent users through pages where no meaningful decision occurred. The second attempt built the experience one screen at a time with explicit instructions and feedback. That approach was slow because the application is written in Rust without a framework, making even simple agent-driven changes take 10 to 20 minutes, and because the resulting design still needed substantial iteration and edge-case work. Rebuilding the flow through the old process would have taken more than eight additional hours. The successful approach was to use ChatGPT Desktop to create a complete HTML mock application covering multiple surfaces, including a browser, the app, macOS passkey dialogs, email, and pages not tied to a specific surface. A dropdown allowed testing from different starting configurations. The prototype took about three hours of iterative decisions and testing, but made the intended behavior and use cases clear. The author then let the agent continue implementation overnight. The resulting recommendation is to have an AI agent build a working HTML prototype first, test every important situation, and only then implement the product in one pass.