Explore Production Data Safely with ClaudeCode and Bauplan

Exploring an unfamiliar dataset is slow, manual work. You query the schema, sample a few rows, check null rates and cardinality, hunt for anomalies, then write up what you found. In this demo, we hand that entire process to Claude Code and watch it run the exploration on its own.

We start by setting up the environment: Claude Code in the terminal, CLAUDE.md pulled from the Bauplan GitHub repo, and the Bauplan skills installed through the Anthropic marketplace and enabled with the plugins command. We then give the agent one instruction, only read, and point it at a single table created in a previous safe ingestion session.

From there the agent works through four phases. It begins with namespace discovery to find what exists. It retrieves the schema and samples rows, then infers what the data represents, in this case one row per show-watch interaction with show metadata and user info. It profiles the table by computing counts, min and max, null rates, and cardinality across every column to surface anomalies. Finally it compiles everything into a written markdown summary.

The whole run is built on a data explorer Python file that Claude writes and updates as it goes, querying through the Bauplan SDK and profiling with Polars. It hits a few errors along the way, a naming issue and several in the Polars code, and each time it proposes a fix, we approve, and it reruns. The result is two deliverables: a reusable explorer script covering all four phases, and a structured report with table semantics, data quality observations, and join candidates.

Because every step is a read, this is a workflow you can point at the tables you have in production without risk of a write touching anything. You get the analysis an engineer would otherwise assemble by hand, plus a script you can rerun on the next dataset.