How To Design Better AI Apps

TL;DR

Current AI apps fail because developers hide system prompts and treat AI as a black box, producing generic outputs that don’t match individual users. The solution is to let users see and edit system prompts, and to give agents tools (e.g., label, archive, draft) to automate repetitive workflows.

Key insights

• The Gmail Gemini draft feature creates emails that don’t sound like the user and requires prompts as long as the output, making AI feel like a chore.
• The root cause is treating the system prompt as hidden code (developer-only) instead of allowing users to edit it to reflect their personal style and rules.
• The “AI horseless carriage” phenomenon: early designs simply replace the old interface (horse) with AI (engine) without rethinking the architecture—e.g., Gmail slots AI into a human-centric app rather