Fixing Data Pipelines with Bauplan and Claude Code

In this demo, we debug and fix a broken data pipeline end to end using Bauplan and Claude Code.

The setup is simple: Claude Code running in the terminal, a Claude.md pulled from Bauplan's GitHub repo and a pipeline built in an earlier session that breaks things downstream on purpose. The Bauplan skill is installed as a plugin through the Anthropic marketplace, so Claude Code has a structured playbook for debugging and fixing data pipelines.

We start in the Bauplan UI, where a failed job shows up with a DAG overview and details on what went wrong. From there, we hand Claude Code a query asking it to look at the latest pipeline run on the branch and it takes over.

What happens in the walkthrough:-

• Claude Code creates a debug folder and works through a step-by-step process
• it finds the latest failed job, checks out the right branch, and writes a job report with the failure metadata
• it creates a Bauplan debug branch to work in- it collects evidence on the downstream table and captures the context in a markdown file
• it locates the root cause: schema drift, where an upstream model renamed a column and the downstream content metrics model still expected the old name
• it applies a minimal fix to the models.py, including a Polars change
• it runs a dry run locally, then re-runs the full pipeline to verify the fix
• all three models execute cleanly and the job completes

To close, Claude Code writes a debug summary report covering the root cause, what broke, why it broke, the fix and the rerun result. We then run the fixed pipeline manually to confirm it and check the completed job back in the Bauplan UI.

The result is a clear division of labor: the agent handles the mechanical debugging steps, and you approve the decisions along the way.