You're building a sales channel analysis for a furniture workshop in Cape Town — computing revenue, margins, and production efficiency across three sales channels, then writing the findings into a client-ready narrative.
The discipline skills: loading data into DuckDB, profiling it against a data dictionary, computing metrics from provided definitions, building charts that answer specific business questions, and writing findings into an executive summary with recommendations.
The AI-direction lesson: this is your first time directing AI through a full analytics workflow. AI will compute the numbers, generate the charts, and draft the findings. Most of it will look right. The skill is noticing when it isn't — when AI computes "total revenue" using a formula that includes refunds the definition excludes, or generates a chart with a truncated y-axis that misrepresents the trend. Checking AI's output against provided verification targets is the practice that makes everything else trustworthy.
Your Role
You're picking up this analysis. You have Claude Code, DuckDB, Jupyter, and the full analytics toolkit from setup. Your job is to direct AI through the work — loading the data, computing the numbers, building the charts, writing the findings — and make sure what comes back is actually right.
AI handles the computation. You handle the thinking. When AI computes "total revenue," it doesn't check whether that number matches what Naledi's business actually means by revenue. That's on you.
What's New
Setup gave you the tools and a single-prompt demo. This is the first time you're doing it for real — for a client who needs the answer.
Everything is provided: an analysis spec that tells you what to compute, metric definitions that tell you what the numbers mean, a data dictionary that documents every column, chart specs, a narrative template, and verification targets so you can check AI's output against known-correct values. Your focus is on the loop itself — understanding the sequence from data to insight to communication — not on figuring out what to do next.
The hard part is not the computation. It's noticing when AI's number looks right but isn't — because it included refunds the definition excludes, or aggregated before deduplicating, or generated a chart with a truncated y-axis that misrepresents the trend.
Tools
- Python 3.11+ (via Miniconda, "analytics" environment)
- DuckDB
- Jupyter Notebook
- pandas
- matplotlib / seaborn
- Claude Code
- Git / GitHub
Materials
You'll work with these provided files:
- Dataset — twelve months of furniture sales data (~850 rows)
- Data dictionary — column definitions, types, constraints, and what each value means
- Analysis spec — the four business questions, expected outputs, chart specifications, and verification targets
- Metric definitions — precise definitions for revenue, profit margin, and revenue per workshop-week
- Narrative template — executive summary, key findings, and recommendation structure