Early attempts by Notion AI’s engineers to leverage large language models (LLMs) involved complicated code generation and intricate data structures. However, through trial and error, the team moved away from complex modeling towards simple prompts, markdown formats, and clear, human-readable instructions. This shift drastically enhanced AI performance, culminating in the September release of Notion’s V3 productivity software, which features customizable AI agents hailed as a “step function improvement.” Ryan Nystrom, AI engineering lead, emphasizes designing prompts as if explaining to a person without context, which helps the AI understand better. Notion also limits the information fed into its AI to avoid performance degradation, focusing on a streamlined set of tools rather than overwhelming users with options. Their philosophy centers on straightforward use of APIs and plain English, embodying a sleek, efficient AI integration that feels intuitive and indispensable.
Back
