Tool Use and Structured Output · ai-skills

Tools, Workflows and Agents · Beginner to Intermediate

A model with tools cannot act directly. It proposes a tool call as structured data, and a program checks the call and runs the tool.

Role	Does	Never does
Model	Reads the task and proposes one tool call as structured data.	Runs code, touches files, or drives pins itself.
Program	Validates the call against the schema and executes the tool.	Trusts proposed text without checks.
Tool	Performs one named action and returns an observation.	Interprets language or chooses what happens next.

Loop from task to model to program to tool and back to the model.

The tool loop: the model proposes, the program validates, the tool executes, and the observation returns as input.

Lesson	What it covers
Describing tools to a model	Names, purposes, and parameter schemas.
Propose, execute, validate	The call loop and argument validation.
Handling malformed calls	Sorting bad calls, rejecting, and asking again.
Idempotent and side-effecting tools	Safe retries, idempotency keys, and approval.
Tool-call validator lab	A runnable Python validator with an eleven-case suite.

Check yourself The model emits the text read_register with addr 8. What actually runs the read?

Nothing in the model runs. The program parses the proposed call, validates it against the schema, and runs the registered tool.

Role

Does

Never does

Model

Reads the task and proposes one tool call as structured data.

Runs code, touches files, or drives pins itself.

Program

Validates the call against the schema and executes the tool.

Trusts proposed text without checks.

Tool

Performs one named action and returns an observation.

Interprets language or chooses what happens next.

Lesson

What it covers

Names, purposes, and parameter schemas.

The call loop and argument validation.

Sorting bad calls, rejecting, and asking again.

Safe retries, idempotency keys, and approval.

A runnable Python validator with an eleven-case suite.