Interview, Fireside Chat
How To Design Better AI Apps
- Core Thesis: Current AI applications in software are often "horseless carriages"—legacy interfaces (like standard email clients) retrofitted with AI engines rather than being redesigned from the ground up to leverage AI's unique capabilities.
- Problem Identification: The current default approach treats AI agents as generic "lowest common denominator" tools, resulting in outputs that feel robotic, impersonal, and often require more effort to correct than doing the work manually.
- Gmail Case Study: Pete Kuhman cites the Gmail/Gemini draft agent as a primary failure of this approach, where the AI output was a formal, generic email that would likely cause the recipient to suspect the sender's account was hacked.
- The "System Prompt" Bottleneck: The generic tone of AI outputs stems from hidden, hardcoded "system prompts" (e.g., "use a businessy tone," "avoid embarrassment") that developers force upon all users to mitigate liability, preventing personalization.
- Proposed Solution: Developers should allow users to view and edit their own system prompts, effectively letting users "program" the AI agent in natural language to reflect their specific workflows, tone, and preferences.
- Analogy: Just as early cars retained carriage designs until the chassis was re-engineered for the engine, AI apps must move beyond simple text generation to "agent" models that can execute complex, multi-step tasks across different software environments.
- User Capability: Kuhman argues that "prompt engineering" is intuitive for non-technical users, similar to how computer literacy became standard, though initial friction will exist for most consumers.
- Future UI Paradigm: Rather than users manually writing prompts, future systems will likely auto-generate and iteratively update system prompts by analyzing user history (e.g., past emails) and feedback, requiring user intervention only in "break glass" scenarios.
- The "Tools" Shift: The real value of AI lies not in chat interfaces but in "tools"—APIs that allow agents to perform actions (labeling, archiving, creating BCCs, interacting with Slack/Calendar) rather than just generating text.
- YC Implementation: Y Combinator has already built internal agents for finance and legal teams that automate repetitive workflows, demonstrating that "AI-native" tools look radically different from traditional software.
- Forward-Looking Statement: Kuhman predicts that within five years, 99% of users will have customized, auto-generated system prompts that they do not need to touch, while the remaining 1% will use these tools to offload all "busy work" from their cognitive load.
- Advice for Founders: Founders should not attempt to "insert" AI into existing products but should redesign tools from scratch to maximize automated task completion, moving away from the "chatbot" paradigm toward "action-oriented" agents.