Core concept: The Claude API has no special "tool" role. A tool result travels back to Claude inside a user message that carries a tool_result block. That block is matched to the original request by its ID.

Why A and B are correct:

The docs say to continue the conversation "by sending a new message with the role of user, and a content block containing the tool_result type".
The block must carry tool_use_id: "The id of the tool use request this is a result for" — the ID Claude generated on its tool_use block.
Ordering also matters: "Tool result blocks must immediately follow their corresponding tool use blocks", and they must come first in that user message's content array.

Why the others are wrong:

C: Assistant messages hold Claude's own output, including tool_use. "user messages include client content and tool_result." is_error is optional and is set to true only when the tool failed.
D: Content may be "a string (for example, "content": "15 degrees")", nested content blocks, or document blocks. A raw string is valid and ends nothing; the loop ends when stop_reason is end_turn.

Remember: Tool results go back as a user message, keyed by tool_use_id.

Question 183