Thariq Shihipar | Claude Code, Anthropic

Thariq Shihipar, from Claude Code at Anthropic, presents a lecture on building mental models for interacting with AI agents like Claude. He begins by discussing the shift from human-computer interaction to human-agent interaction, highlighting the complexity of AI models as 'grown, not designed' entities that exhibit emergent properties. Shihipar elaborates on the concept of 'capability overhang,' where models possess more capabilities than initially apparent, and how understanding this requires iterative experimentation and a grasp of the underlying prompts and processes. He introduces practical frameworks for staying in the loop with AI development, such as exploration, spec interviews, technical plans, implementation notes, and explainers. Shihipar uses a case study about Pokémon names ending in 'aw' to illustrate how Claude can execute complex tasks when prompted correctly, contrasting this with a simpler, less informative prompt. He concludes by emphasizing the importance of understanding the multifaceted relationship between the agent, the model, the harness (tools and processes), and the world, which ultimately defines the agent's utility and success.