What they did
Thu Vu, a data analyst and YouTuber, wanted AI help on a table of 2,000 hotel guest records with emails, billing addresses and card numbers, without the real rows ever leaving the laptop. Vu first used the Synthetic Data Vault Python library locally to learn the table's patterns and sample a 100-row fake version with the same columns and types. Vu then opened Claude Code in cloud mode, so it had no access to local files, attached only the fake CSV, and asked for a Jupyter notebook using pandas and Plotly that answers three questions: typical length of stay, most common email domains, and how many guests are households versus corporate accounts. The prompt also asked for comments on data gotchas and a live dashboard. The guest data in the demo is itself made up.
How it went
Claude Code wrote and tested the notebook against the synthetic file, including data quality checks, length-of-stay buckets, an email domain breakdown and the household versus corporate split, and built a preview dashboard. When the downloaded notebook ran locally, Plotly failed on a missing nbformat package, so a second prompt got a requirements.txt. After swapping the file path to the full dataset, the notebook ran locally. The code, both prompts and the synthetic data workflow are published in a repo alongside a video.
Worth knowing
The point is that the agent only needs the shape of the data to write the code, not the data itself. Synthetic generation can misread columns: SDV detected the billing address column as an id and it had to be corrected by hand. The author also advises dropping columns you do not need, such as card numbers, before uploading even the fake copy.