How To Design Better AI Apps

YC partner Pete Koomen argues most AI apps are horseless carriages — bolting AI onto legacy UX instead of redesigning around user-programmable automation.

• Gmail’s Gemini draft agent fails because its system prompt is hidden, generic, and corporate-safe — producing emails no real user would write.
• Exposing the system prompt and letting Pete write his own (43yo dad, keep emails short) cut a sick-day email to 2 sentences.
• The horseless carriage problem: AI teams ask ‘how do I slot AI in?’ instead of ‘how do I redesign this to offload repetitive work?’
• Coding agents (Cursor, Windsurf) feel magical because they give bare-metal access and don’t nerf the model for liability reasons.
• Liability shift argument: if users control the system prompt, embarrassing outputs are on the user, not Google — removing the need to sanitize outputs.
• YC is already running internal agents where finance/legal staff iterate system prompts collaboratively, not by editing raw text.
• YC-batch company Den is building ‘Cursor for knowledge work’ — chaining MCP tool servers so one agent spans Slack, Google Docs, email, GitHub, and Linear.
• Pete’s forecast: in 5 years, 99% of users won’t manually edit system prompts — the AI will auto-generate and update them from usage history and feedback.

2025-05-23 · Watch on YouTube