explore

Skill: Explore Data

Purpose

Quick, interactive data exploration without the full pipeline. Lets users poke around the active dataset — preview tables, check distributions, spot patterns, and form hypotheses before committing to a formal analysis.

When to Use

• User says /explore or "let me explore the data" or "what's in this dataset?"
• After connecting a new dataset, before any formal analysis
• When the user wants to understand data shape without a specific question

Invocation

/explore — explore the active dataset /explore {table} — focus on a specific table /explore {table} {column} — deep-dive into a specific column

Instructions

Step 1: Load Context (with Fallback)

Try to load formal context first:

1. Check if .knowledge/active.yaml exists
2. If yes, read it to identify the active dataset name
3. Read .knowledge/datasets/{active}/schema.md for table/column reference
4. Read .knowledge/datasets/{active}/quirks.md for known gotchas

If .knowledge/ files don't exist (common for new users), fall back:

1. Look for data in these locations (in order):
   • data/examples/*.csv (shared example datasets)
   • data/practice/*.csv (if a local practice dataset is present)
   • tests/fixtures/*.csv (test data)
2. Use the first location where data is found
3. Infer schema by reading a sample of the data
4. Proceed with exploration using discovered data

If no data found anywhere:

• Prompt: "No dataset found. Use /connect-data to add one, or point me to your data files."
• Do NOT proceed with hypothetical exploration

Step 2: Choose Exploration Mode

Mode A: Dataset overview (no table specified) — READ-AND-STEER

Goal: Give the user just enough orientation to steer, then stop and ask what they want to explore. Do NOT autopilot into observations, findings, or starting questions. The tool's job is to read the situation and hand the steering wheel back to the user.

Deliver (keep it tight — this is a one-screen opener, not an analysis):

1. Dataset identity: name, source/connection, coverage window.
2. Table list: names with row counts. One-line format per table, no embellishment.
3. Entity map: a short diagram of how the tables relate (e.g., users → orders → order_items → products).
4. Stop and ask. End with an open question, e.g. "What would you like to explore in {dataset}?"