Structured data with tools

**Structured Output: Prompt vs. Tools**

• Two Methods for Structured Output:
• Prompt-Based: Uses techniques like message pre-fill and stop sequences to force Claude to output JSON. (Easier to set up. )
• Tool-Based: Uses defined tools and schemas to reliably extract structured data (primarily JSON). (More complex to set up, but more reliable. )
• When to Use Each Method: Both methods are valuable; choose prompt-based for simplicity or tool-based for maximum reliability.

Tool-Based Structured Output Process

• Core Concept: Define a tool whose input parameters dictate the required data structure.
• General Steps:
• Define Schema: Write a JSON schema specification for a tool. The inputs of this schema define the structure of the data you want to extract (e. g. , balance must be an integer, key insights must be a list of strings).
• Input to Claude: Feed the JSON schema specification and the raw data source into Claude.
• Claude's Action: Claude identifies the tool and responds with a tool use message containing the extracted data as arguments (the desired JSON).
• Completion: The process is complete once the structured data is returned; no further conversation is needed.

Implementing Data Extraction with Tools

• Goal: Extract specific data (e. g. , title, author, topics) from a text source.
• Steps:
• Create Tool Schema: Define a tool (e. g. , article details schema) specifying the required inputs (e. g. , title → string, author → string, topics → list of strings).
• User Message: Write a prompt asking Claude to analyze the text and call the defined tool.
• Tool Choice (Forcing Use): To ensure Claude uses the tool, set the tool_choice parameter in the request (e. g. , set it to the specific tool name, article details tool).
• Result: Claude executes the tool call, and the output is the extracted data in the required JSON format.

Takeaways

• Structured output can be achieved using Prompt-Based methods (simpler) or Tool-Based methods (more reliable).
• Tool-Based extraction relies on defining a JSON schema that specifies the required data structure for a tool.
• The process involves feeding the JSON schema and the raw data to Claude, which then responds with a tool use message containing the extracted data.
• To ensure Claude performs the extraction, the tool_choice parameter must be set to force the use of the defined tool.