AI summary by TwinMind
How To Design Better AI Apps
Y Combinator
30m 8s
Introduction and Background
- Pete Kumman, YC partner and Optimizely founder, discusses essay criticizing current AI agent development
- Contrasts two AI experiences: powerful tools like Cursor/Windsurf vs frustrating existing app integrations
- Gmail's Gemini integration used as primary example of poor AI implementation
Gmail AI Integration Problems
- Gmail's draft writing agent produces formal, generic emails that don't match user's voice
- Example prompt: "Let my boss Gary know my daughter has flu, can't come to office"
- Output: Overly formal "Dear Gary" email that sounds like account was hacked
- Two core problems:
- Doesn't sound like actual user
- Prompt length roughly equals draft length (inefficient)
System Prompts and Hidden Design Decisions
- Gmail combines user prompt with hidden "system prompt"
- System prompt defines AI's role, tone, behavior - reused for all users
- Estimated Gmail system prompt: formal, businessy tone, one-size-fits-all approach
- Google prioritizes safety over personalization to avoid embarrassment
- Demo shows impact: generic system prompt vs personalized "You're Pete" prompt
- Personalized version produces: "Hi Gary, my daughter's sick with flu, can't come in today. Thanks"
The AI Horseless Carriage Problem
- Reference to early automobiles that looked like carriages with engines
- Current AI development uses old software paradigms instead of leveraging AI capabilities
- Traditional software: developer writes code, user uses interface, one-size-fits-all approach
- AI enables user-programmable software through natural language
- Gmail team asked "How do we slot AI into Gmail?" instead of rethinking from scratch
Why Coding Agents Excel
- Two reasons coding agents (Cursor, Windsurf) work better:
- AI models excellent at processing text instructions into code output
- Developer tools are "power tools" - allow full access without safety restrictions
- Other domains still use "kid glove mentality" limiting AI capabilities
- Liability concerns prevent companies from giving users full AI access
User Programmability and System Prompt Access
- Argues users should see and edit system prompts
- Shifts responsibility from company to user (like Gmail email content responsibility)
- Most people can't write system prompts today but will learn quickly
- Comparison to computer adoption: once seen as "nerd tools," now universal
- Prompting more accessible than traditional computing - just requires English explanation
Practical Implementation Strategies
- Users shouldn't write prompts from scratch
- AI should analyze 20 years of email history to create initial draft prompts
- Human assistant analogy: gradual training through interaction, not 30-page instruction manual
- Iterative process: AI writes, user provides feedback, system auto-updates
- Need "AI system prompt writer" tool to translate user feedback into prompt updates
YC Internal Tools Experience
- Pete built internal AI tools for YC finance/legal teams
- Process: sit with team members, understand workflows, write system prompts together
- Iterative refinement through user feedback and prompt editing
- Demonstrates feasibility of user-customized AI agents
Tools and Agent Capabilities
- System prompts define behavior; tools enable action
- Email agent example tools: label emails, archive, write drafts
- Vision: agents handle transactional work (bills, introductions, handoffs)
- Email inbox as to-do list of chores, not creative work
- Integration across platforms: Slack, calendar, Notion, Jira, Linear
- Example company: Den (YC batch) building "cursor for knowledge work"
- MCP servers enable tool chaining across applications
Future Vision and Founder Advice
- Steve Jobs: software as "bicycle for the mind"
- AI agents: "rocket ship for the mind"
- Move beyond chatbot paradigm to work automation
- For founders: rethink tools from ground up with AI, don't just embed chatbots
- Ask: "How would I design this tool to offload repetitive work?" not "How do I insert AI?"
- Opportunity to rebuild every existing tool with AI-native approach
