Let AI explore your data before you ask it anything

Executive overview

Most people constrain AI by opening with a specific question — the AI focuses only on skills relevant to that question and misses everything else. Share your data and intent instead, then get out of the way.

The fix is a three-phase process: explore first, discover patterns, then ask targeted questions only about surprises.

Letting AI explore before you question it surfaces insights you didn't know to look for.

The problem with asking questions first

• A targeted question forces the AI to use only the skills relevant to that query.
• This cuts off its ability to combine domain knowledge with data analysis.
• You only get answers to questions you already knew to ask.
• Novel insights — the ones that change decisions — come from questions you didn't know to ask.

The three-phase approach

• Phase 1 — Explore: give the AI your data and your intent, not a question.
• Phase 2 — Discover: the AI surfaces patterns, anomalies, correlations, and surprises.
• Phase 3 — Constrain: pick the surprising findings and ask specific follow-up questions only then.

Method 1: basic chat app (ChatGPT, Claude, Gemini)

• Paste a structured prompt, attach the data file(s), and let it run.
• Prompt structure: role the AI should play → your role and intent → instruction to explore with fresh eyes → ask for what it found, why it matters, what to do → ask it to go beyond the obvious.
• After getting findings, ask targeted questions only on the two or three surprises.
• File size limits apply: CSVs are lightest; Excel with multiple tabs is heavy; complex PDFs consume the most context.
• Cap is roughly 10 files per conversation.
• When using Excel, export the relevant tab as a CSV before passing it to the AI.

Method 2: Claude Code or Codex (AI comes to the data)

• Create a single folder on your machine, name it after the analysis, and dump all files into it.
• Add a claude.md (or agents.md for Codex) file with the task and rules — this is the AI's system prompt.
• No file quantity or size limits; the AI externalises its memory rather than loading everything into context.
• The task section mirrors Method 1: industry, data type, your role, your intent, explore every file, produce a ranked findings report.
• Rules section handles memory resets — the key problem when processing hundreds of files.

Handling memory resets in Method 2

• On each memory reset, the AI re-reads claude.md to recover its task and rules.
• It checks a todos file to see what has already been processed.
• If todos doesn't exist yet, it creates it first, listing every file to process.
• Each file gets checked off as it is processed.
• An insights file accumulates findings continuously across resets.
• This loop — read instructions → check todos → process file → write insights → check off → repeat — runs until all files are done.

File format guidance for Method 1

• CSV: lightest, AI spends more time reasoning than reading.
• Markdown and TXT: slightly heavier, still manageable.
• Excel / Google Sheets (multi-tab): can be overwhelming; export individual tabs as CSV.
• PDF: heaviest if large or complex (screenshots, diagrams, annotations).