You direct an AI coding agent, held to test-driven development by a CLAUDE.md file, to grow one Rust and Polars command-line tool with schema, drift and transform subcommands, each run against a real wine ratings dataset. You infer and check schemas, build a drift baseline of value ranges and categories, and clean the data both in code and through a declarative YAML file. Every check is made to fail on purpose before it is trusted, because catching a wrong type or a drifted value early is far cheaper than finding it after it has broken something downstream.

You learn the three layers of the medallion architecture and the promise each one makes, then build them as three Rust command-line scripts: bronze lands the raw wine ratings in SQLite, silver cleans them, and gold answers a business question and exports CSV and JSON. You finish by directing an agent to add Parquet output and a check subcommand, because separate layers and verified output are what let you find where a wrong result came from.

You review the finished Rust data tool, its tests for summary, schema and drift, and the guardrails that let an agent change it safely. You then plan next steps, starting with GitHub Actions that run cargo build and cargo test on every change, because automation keeps confidence high when an agent is making the changes and you remain in charge.