BlogSource
Coding agents for data analysis
by Simon Willison · 2026-03-16
✓ Opened with WebFetch on 2026-10-15
Read the original ↗Part 1 · What the source did
A workshop handout for data journalists: use Claude Code or Codex to ask questions of a database, clean messy codes, build visualizations and scrape data.
- Ran a 3-hour workshop on GitHub Codespaces with a budget-restricted Codex key.
- Had attendees query SQLite, decode neighborhood codes and scrape with agents.
- Served a viz/ folder through Datasette and had Claude Code build an interactive Leaflet heat map there.
Summary written by Agent Cookbook. See the original for the full context.
Part 2 · Our copy-paste version
Adapted by us from what the source described. Fill in the [BRACKETS].
The prompt
You have [DATABASE.db] (SQLite). Work only in this folder.
1. List the tables and describe each in one line. Show 5 sample rows from the most important one.
2. Answer: [QUESTION]. Show the SQL you ran and the result.
3. Find columns with codes or messy values and propose a cleaning plan. Don't modify the original; write a cleaned copy.
4. Build one static HTML visualization in viz/ that answers the question. Use a CDN library, no build step.
🛡
Approval ruleThe agent writes to a copy; the original data file stays read-only.
Set it up
- Put the data in a fresh folder or Codespace.
- Set a spending cap on the API key if you use one.
- Run the prompt step by step and check the SQL yourself.
- Open the viz file locally.
Common failures
⚠
What usually goes wrong
- Agents state numbers confidently; always ask for the query behind a figure.
- Scraping steps may break site terms. Check before running.
Variations
- Point Datasette at viz/ so new charts show up live.
- Use the same flow for spreadsheets exported to CSV.