Tool Use Examples is the name Anthropic gives to a feature of the Claude Developer
Platform that lets a developer attach sample tool calls to a tool definition, in an input_examples
field alongside the input schema. Its premise is that JSON Schema defines structure — types, required
fields, allowed values — but cannot express usage patterns: which format a date field should take,
what an identifier looks like, when to populate a nested object, or how one parameter relates to
another. The examples show those patterns concretely, so that the model can infer them rather than
guess, with the aim of reducing malformed tool calls and inconsistent parameter use.
Usage
The feature was introduced in Introducing advanced tool use on the Claude Developer Platform as one of three beta features for agents working with large tool libraries, alongside the Tool Search Tool (Tool search) and Programmatic tool calling, and is positioned as the answer to the third of their bottlenecks: parameter errors and malformed calls. The announcement illustrates it with a support-ticket tool whose three examples — a fully specified critical bug, a feature request with a reporter but no contact or escalation details, and a task with only a title — are meant to teach date and ID formats, how to build the nested reporter object, and which optional parameters appear together at each level of priority.
When It Applies
Anthropic describes the examples as most useful for complex nested structures where valid JSON does not imply correct usage, tools with many optional parameters whose inclusion matters, APIs with domain-specific conventions not captured in the schema, and similar tools that examples help tell apart. It describes them as less useful for simple single-parameter tools, standard formats such as URLs or email addresses that the model already understands, and validation that JSON Schema constraints handle better. Because examples add tokens to every tool definition, the stated trade-off is whether the accuracy gain outweighs that cost.
Its guidance on writing them is to use realistic data rather than placeholder values, to show variety across minimal, partial and full specifications, to keep to one to five examples per tool, and to add examples only where correct usage is not obvious from the schema. The evidence behind the feature is Anthropic’s own: the announcement reports that in internal testing, tool use examples improved accuracy on complex parameter handling from 72% to 90%.