Sending tool results

Step 4: Running the Tool Function

• Access the tool use block from Claude's response via response. content[1]
• Extract the input arguments (dictionary) using . input property
• Convert the dictionary to keyword arguments using the ** operator when calling the function
• Example: get_current_datetime(**response. content[1]. input)

Step 5: Sending Follow-up Request with Tool Results

Tool Result Block Structure

• Type: tool_result (placed inside a user message)
• Purpose: Feeds the tool function's output back to Claude
• Key properties:
• tool_use_id: Matches the ID from the original tool use block to link request with result
• content: The output from the tool function (converted to string/JSON)
• is_error: Boolean flag (default false) indicating if the tool execution failed

Why Tool Use IDs Matter

• Claude may request multiple tool calls in one response
• Each tool use block gets a unique ID
• Tool result blocks must reference the correct ID to match results with requests
• Prevents ambiguity when multiple tools are called

Complete Conversation Flow

• Send user message + tool schema to Claude
• Claude responds with assistant message containing tool use block(s)
• Execute the tool function(s) locally
• Send follow-up request with:
• Full conversation history (original user message + assistant response)
• New user message containing tool result block(s)
• Original tool schema(s) again
• Claude processes the tool result and sends final response

Takeaways

• Extract tool inputs from Claude's response using . input property and call the function with ** operator to unpack arguments
• Tool result blocks must include tool_use_id to match results with the correct tool call when multiple tools are invoked
• Follow-up requests require the full conversation history, tool result block(s) in a user message, and the original tool schema(s) for Claude to process results correctly