How To Design Better AI Apps

Kumman contrasts coding tools like Cursor and Windsurf, which help users create what they imagine, with AI features in familiar apps that can feel like extra work. In Gmail, a draft agent turns a short request about his sick daughter into a formal email that does not sound like him. The prompt is nearly as long as the draft.

Kumman says generic system prompts can produce safe, one-size-fits-all results. He demonstrates that a prompt describing his role and preference for short emails produces a draft closer to his own style. He argues that users should be able to inspect and adjust these instructions.

Kumman uses the “horseless carriage” analogy for adding AI to existing software without redesigning it around AI’s capabilities. He compares this to early search engines that were digital directories and mobile apps that were websites in native wrappers. For email, he proposes an agent that can label, archive, and draft messages according to personal instructions.

Kumman describes AI models as capable assistants that need users to explain what work to do. He suggests agents could learn from past work and feedback, while users retain access to the instructions when they need to correct them. With tools connected across email, Slack, calendars, and documents, agents could handle repetitive workflows across apps.

Kumman advises founders to rethink existing tools from the ground up instead of simply inserting a chatbot. He says product teams should ask how AI can take repetitive work off users’ hands so they can focus on what matters.