Ask, Then Edit

You'll learn: the Ask tab — Claude answering data questions by writing analysis programs — and the round-trip that makes it more than a chatbot: every AI-run program lands in the Graph tab where you can inspect it, change it, and re-run it, with the conversation kept in the loop.

The three tabs are three authors for the same language. Ask is the fastest first draft; the Graph tab is where you make the analysis yours.

Set up the Ask tab#

1. Open ⚗ Analysis → Ask. First visit wants an Anthropic API key: paste one and press Save key. It's stored in this browser's localStorage and sent only to the Anthropic API — the data never goes anywhere. Programs run locally in your page, exactly like preset and graph runs; the model only sees the program it wrote and the summary statistics that came back (r values, means, coverage — a few hundred bytes).
2. The model gets a system prompt generated from the live layer catalog, so it knows what data exists, the coverage windows, and the op set — and its run_analysis tool schema enumerates the real language (op names and layer keys as enums). If a program is still wrong in any way, the engine's validator sends the structured errors back and the model repairs it in the same turn.

Ask#

3. Type: "Is there a correlation between sea temperature and wind?"

   Watch the sequence: a thinking note, then a collapsible ▶ ran analysis program (5 nodes) line — expand it and you'll recognize everything from tutorial 2: two layer nodes, a temporal correlate, a display, an answer. The map takes over with the r map, the results area fills in, and the model narrates the actual numbers from the run — it can't make them up, because the only numbers it has are the ones the engine returned.

4. If the model ever emits an invalid program, the same structured validation errors you've seen as red badges go back to it as the tool result, and it repairs the program and retries within the same turn. You'll see two program expanders in that case — the broken draft and the fix.

Steal the program#

5. Switch to the Graph tab: the model's program is already loaded (the Graph tab always shows the most recent run, whoever authored it). This is the "show your work" view — every claim in the chat corresponds to a node you can poke.
6. Edit it. Change the correlate window (start on both layers), or insert a mask with a region node to focus the tropics. Press ▶ Run. The map and answers update — no model call, no tokens, sub-second.
7. Now go back to Ask and follow up: "what about only in the tropics?" Your manual run was appended to the conversation (as a note carrying the edited program), so the model knows what's currently on screen and edits from it rather than starting over — usually changing a node or two, which you can verify in the next program expander.

Keep the good ones#

8. When an AI-drafted analysis is worth keeping, save it from the Graph tab (💾 with a name), or 🔗 Copy link to share the exact program as a URL. Links and presets are plain program JSON — nothing about them remembers which tab authored them.

Which tab when#

One habit to build: whatever the author — model, preset, or you — the result always arrives as map + numbers + the program that made them. Reading the program is how you trust the answer.

---

That's the series. The module guide is the complete node-type reference; the book's Geo GIS Explorer deep-dive covers the explorer around the analysis engine — the data feeds, the time-lapse machinery, and how the display takeover works.