TOOL FORGE
A tool use & function calling dojo
0 / 900 XP
Anatomy of a reliable tool call

Nine belts. Real tool calls, real judgment.

Every belt drills one core skill behind reliable tool-using assistants: when to call a tool at all, how to define it precisely, how to handle it failing, and how to know when an agent should stop. You write the system instruction, Claude actually runs it against a live user request, and your Coach grades the instruction itself against the standard.

White → Brown: the fundamentals — judgment, schemas, disambiguation, errors, grounding. Black & 1st Dan: full multi-tool systems and agentic stopping conditions.

{"name": "search_orders", "description": "Look up a customer's past orders by email and status.", "parameters": {"customer_email": "string, required", "status": "enum: pending|shipped|delivered|canceled"}, "max_results": "integer, 1-50, required"}
Name & descriptionWhat the tool does, in words the model can act on.
Constrained parametersEnums and types, not freeform strings.
Required vs optionalNo ambiguity about what must be supplied.
Bounded rangesLimits that make malformed calls impossible.