docs.claude.comAnthropic docs
Framework / Claude
Claude is the reasoning core at the center of every autonomous system I build. This section covers the model variants, how to prompt for agents, extended thinking, and the mechanics of tool use, everything you need to make the model do reliable work.
The Claude family, Opus, Sonnet, Haiku, when to reach for each, and how to think about capability versus cost.
How prompting changes when the model is running in a loop with tools. System prompts, roles, tool-call contracts.
When to turn on reasoning mode, how it changes outputs, and the trade-off with latency and cost.
The mechanics of tool calls, argument schemas, error recovery, and the patterns that keep multi-step agents on track.
Andrej Karpathy - State of GPT (Microsoft Build)