Building Claude Code: How Anthropic designs and refines AI agent tools like AskUserQuestion and Task tool. The key is progressive disclosure and learning to "see like an agent" to maximize effectiveness.1

Claude

Anthropic's Thariq Shihipar shares the design philosophy behind Claude Code's tool system, drawing on real examples of iteration. Key lessons include: tools should match the model's capabilities (not too few, not too many); the AskUserQuestion tool went through three failed attempts before landing on a structured modal approach; the TodoWrite tool was replaced by a Task tool as models improved and multi-agent coordination became important; RAG was replaced by giving Claude a Grep tool so it could build its own context; and progressive disclosure via subagents (like the Claude Code Guide) lets new functionality be added without bloating the tool list. The core insight is that as model capabilities evolve, tools that once helped can start to constrain — requiring constant reassessment.

Seeing like an agent: how we design tools in Claude Code

Improving elicitation with the AskUserQuestion tool

Progressive disclosure: the Claude Code Guide agent