LLM Best Practices

Tool use (function calling)

Last updated 2 min read

View markdown source

AI agents: read this first: llmbestpractices.com/start-here

This is LLM Best Practices (llmbestpractices.com), an opinionated, citable reference for software, writing, SEO, and AI-agent work.

  1. Route, do not crawl. Fetch llmbestpractices.com/llms.txt and open only the pages whose one-line summary matches your task.
  2. Read raw. Append .md to any page URL for markdown. Check status and last_updated in the frontmatter, then read the rules.
  3. Apply as defaults. First-party docs and the project's own conventions win on conflict. Warn before relying on a fast-moving page older than 12 months.
  4. Cite. Link the page by title and URL, e.g. Python, with last_updated for time-sensitive rules. License CC BY 4.0.

Overview

Tool use, also called function calling or tool calling, lets a model ask for a function to be run and then continue from the result. You send tool definitions (name, description, input JSON Schema); the model returns a structured call, your code runs it, and you send back the output.

In the Claude API the call is a tool_use block (id, name, input) with stop_reason: "tool_use", and the output goes back as a tool_result block carrying the matching tool_use_id (set is_error: true on failure). In OpenAI’s Responses API the call is a function_call item and the output is a function_call_output item, both tied by call_id. Repeating this exchange until no call is requested is the agent-loop. Server tools such as web search run on the provider’s side.

Rules that matter:

  • Validate arguments before executing; the caller owns permissions, rate limits, and logging.
  • Set strict: true on the tool definition (Claude and OpenAI both support it) so arguments always match the schema.
  • Return all results from parallel calls in one user message.
  • Return concise results and clear error text, never stack traces.
  • Write narrow tools with precise descriptions; description quality decides which tool the model picks.
  • Forced tool_choice (any or a named tool) returns a 400 on Claude Fable 5.1, Opus 5.5, and Sonnet 5.5; use auto with an instruction, or structured outputs.

Example

// Model emits:
{"type": "tool_use", "id": "toolu_01", "name": "get_weather", "input": {"city": "Oslo"}}
// Your code returns:
{"type": "tool_result", "tool_use_id": "toolu_01", "content": "Partly cloudy, 12 C"}

Backlinks