Retail Analyst
Turns an executive’s question into a finished analysis: checked numbers, a plain-English answer, and a dashboard you can forward.
Walkthrough video coming soon.
Step by step
An executive’s question comes in, like “How are same-store sales by market?”
The AI confirms the timeframe, markets, and metric before it pulls anything
It reads the standard definitions and checks the known data issues
It writes, runs, and checks commented SQL against the database
A person decides: The analyst reviews the findings and decides what matters and what to tell the executive
It builds a dashboard in the company brand that works offline and can be emailed
Every query, decision, and file is saved in the analysis’s folder in the workbench
What it does
An analyst brings it a question from an executive. It asks what it needs to know, writes and checks the SQL, walks the analyst through what it found, and builds the dashboard once the analyst is happy. The AI does the technical work; the analyst decides what matters and what to say. Every step lands in a workbench folder the analyst can read.
The demo data
Actual Burger Co. is a made-up fast-casual chain: 48 restaurants across the US, Canada, and Mexico, with two years of orders. The data has two real-world problems built in: Mexico can’t separate delivery tips and fees from sales, and its item lines have no prices. Each is documented with a rule for querying around it and what to tell an executive, so you can see how the system handles messy data honestly.
Try it
Open the folder in Claude Code and ask these, in order. Each one shows the next part of the system.
- “What can you do, and where should I start?”
- “Are there any known data issues I should know about?”
- “Write me a query for average check by channel in Canada.”
- “How are same-store sales by market?”
The repo also has a finished example: open its dashboard, then read its record to see every decision behind it.
Use your own data
The data can live anywhere the AI can query, such as BigQuery or Snowflake. To switch, change three things: how the query tool connects, where the data lives, and the data model docs (your tables, rules, and metric definitions).
How it’s built
Context
The business, every table and column, and the standard metric definitions with their SQL. It also documents the known data issues, with what to tell an executive about each one, so the AI never has to guess.
Process
Four skills that follow the same method every time: ad hoc analysis (clarify, pull, check, present, revise, build), a query builder, a dashboard builder, and setup for the first run.
Tools
A local DuckDB database and a read-only query tool (Python, with DuckDB and pandas). Scripts build each dashboard with Chart.js and check it in a real browser before it’s handed over.
Structure
A workbench where every output lands. Each analysis gets its own dated folder: a record of every decision, each query with its data, and the dashboard. Nothing gets lost.
What you’ll need
- Agentic coding tool
- Claude Code. Codex, Gemini CLI, and other agentic coding tools work too, after one copy step the README explains.
- Accounts
- Just the one for your coding tool. The demo data runs on your computer.
- Connections
- None. You need Python 3.9 or newer. On the first run the system sets up Python and downloads the demo database (about 90 MB), asking before each step.