Business :: Ideas :: QuietInbox :: How To Design Better AI Apps

readme.md

Business :: Ideas :: QuietInbox :: How To Design Better AI Apps

⪼ Made with 💜 by Polyglot.

• ..

How To Design Better AI Apps

Summary

This video is a podcast-style interview between Y Combinator partner Pete Kooman and the Breakdown team, with the intent to educate and inspire developers, founders, and AI builders to rethink how AI tools are integrated into products. Kooman critiques current AI integration practices—especially “AI add-ons” like Gmail’s draft assistant—as outdated, restrictive, and underwhelming. He argues that most teams are still designing AI features using old software paradigms that fail to unlock AI's true potential.

Instead, Kooman advocates for AI-native experiences where users are empowered to customize how AI performs tasks—primarily through editable system prompts and agentic tools. He provides demos of AI agents that automate email triage and emphasizes the importance of allowing users to “program” apps with natural language. The conversation explores the potential of agents, editable prompts, and the future of AI as an enabler of personalized automation across domains, not just software development.

Highlights

• Old vs. New Paradigm: Current AI features mimic legacy software development approaches (e.g., one-size-fits-all). True AI-native tools require a shift in design thinking.

• Example Critique – Gmail’s AI Drafting Tool:

  • Generates generic, overly formal emails.
  • Hides the system prompt and restricts user customization.
  • Ends up being more work, not less, for users.

• Editable System Prompts:

  • Should be visible and modifiable by the user.
  • Allow personalization of tone and decision-making logic.
  • Make AI feel like a superpower, not a burden.

• User-Controlled Agents:

  • Kooman showcases an "email reading agent" that applies labels, drafts replies, and triages emails based on user-defined logic.
  • This turns AI into an autonomous assistant, not just a text generator.

• AI as a Platform for Custom Automation:

  • System prompts should evolve into "personal operating procedures" editable in plain language.
  • Future interfaces might help users iteratively refine prompts through conversation or AI coaching.

• Tooling is Critical:

  • Developers should focus on exposing powerful, well-defined tools that AI agents can call (e.g., labeling, replying, archiving).
  • Inspired by systems like Cursor, Windsurf, and Den.

• The “AI Horseless Carriage” Trap:

  • Like early cars that mimicked horse-drawn carriages, early AI features mimic old UIs.
  • Real innovation comes when we design experiences around AI from scratch.

• LLMs Are Not Just Chatbots:

  • The chatbot interface is a stepping stone, not the end goal.
  • LLMs should be used to do work, not just talk about it.

• Power Users vs. General Users:

  • While not everyone will edit prompts manually, AI should be able to auto-generate and iteratively update them.
  • Think: hiring a digital assistant that learns by observing and refining.

• Advice for Founders:

  • Don’t ask how to add AI to your app—ask what your app would look like if built from scratch to leverage AI.
  • Prioritize removing repetitive work from users, not just enhancing interfaces with generative text.

• Final Vision:

  • AI should empower any user to automate routine tasks across tools (email, Slack, Notion, GitHub).
  • Each profession will eventually have its own “Cursor moment” where agents become commonplace.

---