You're investigating why a growing pet food company in Austin, Texas has revenue up 40% but cash tighter than it should be — framing your own analytical questions for the first time, with no analysis spec and no list of questions to answer.
The discipline skills: refining an ambiguous business complaint into specific, testable investigations, time series decomposition, anomaly detection, and delivering the same findings in different formats for different stakeholders (the VP who needs to understand what's broken, the co-founder who needs to know what to do about it).
The AI-direction lesson: directing AI on ambiguous terrain. AI accepts "our numbers feel off" at face value and produces a comprehensive generic exploration that calculates every metric in the dataset and answers nothing specific. The skill is keeping AI focused — decomposing a vague problem into AI-sized analytical pieces, directing each piece with constraints, and resisting the pull toward breadth. Every project before this provided the questions. Now you own them.
Your Role
You're investigating why a growing company's metrics don't match its financial reality. There is no analysis spec, no list of questions, and no investigation template. You decide what to investigate, in what order, and why.
When you find the answer, you deliver it twice: once for the VP and once for his co-founder. Same findings, different formats, different decisions.
What's New
Last time, you assessed experiment validity for a loyalty program, authored your first project memory files, and experienced the difference between AI with definitions and AI without them.
This time, nobody tells you what to investigate. Every project before this one provided the questions -- you directed AI to answer them. Now someone says "our numbers feel off" and you figure out what that means. The question-framing work is yours. AI will happily explore the entire dataset and produce a comprehensive report that answers nothing specific. Your job is to focus the investigation on what matters.
Tools
- Claude Code (with DuckDB MCP active)
- DuckDB (local database, MCP-connected)
- Python (pandas, matplotlib/seaborn, scipy.stats)
- Jupyter notebooks
- Git and GitHub
- CLAUDE.md (project memory -- established practice)
Materials
- Subscription data -- customer-level subscription records covering 18 months.
- Transaction history -- delivery-level transactions with amounts, discounts, and refund flags.
- Marketing spend -- channel-level spend and attribution data with two different attribution models.
- Fulfillment costs -- monthly cost breakdown including packaging and shipping components.
- Data dictionary -- column descriptions for all four data files.
- CLAUDE.md skeleton -- a minimal project governance file. You populate it as the investigation unfolds.