Tech Learning Session — Writing Effective Tools for AI Agents

A short, hands-on assignment based on Anthropic's Writing tools for AI agents.

What is this?

You get one small weather tool (an MCP server) wired into Claude Code. The tool works, but it's deliberately written badly — a vague name, no description, a raw response. Your job is to improve it across 4 rounds and watch the AI agent go from ignoring the tool to using it perfectly — without ever changing what the tool actually does.

The big idea: the agent never sees your code. It only sees the tool's name, parameters, description, and return value. Those four things decide whether it gets called and used well.

Related MCP server: mcp-poc

What you'll practice

• Naming a tool and its parameters so an agent knows when to use it

• Writing a description that is the interface

• Returning readable, useful results instead of raw data

• Writing error messages an agent can recover from

• Keeping responses token-efficient

How to do it

1. SETUP.md — get the weather tool running in Claude Code (~5 min).

2. QUESTION.md — the assignment: what to ask, the 4 rounds, and how to check whether the agent called your tool.

Weather data comes from Open-Meteo — free, global, no API key required.

Requirements

Python ≥ 3.10, Claude Code, and either uv (recommended) or pip. Full details in SETUP.md.