What you'll take away
- Start with a "grill me" session: the agent interviews you, one question at a time, until you both agree on what you're building.
- Turn that into a short product document, then split it into small, separate tasks with the order they have to happen in.
- Keep each task small. Models get worse as their context fills, well before the advertised limit.
- Let the agent work through the tasks on its own, test first, then review its work in a fresh session.
More talks
Talk26mClaude Code best practices
Cal Rueb · Code w/ Claude 2025
One of the people who builds Claude Code at Anthropic on how to get the most out of it day to day.
Workshop25mPrompting 101
Hannah Moran and Christian Ryan · Code w/ Claude 2025
Anthropic's Applied AI team build a real prompt step by step, from a vague first try to one that works reliably.
Talk39mAndrej Karpathy: Software Is Changing (Again)
Andrej Karpathy · YC AI Startup School 2025
How writing software in English changes what you build and how you should design AI tools.
