10 Prompts That Turn ChatGPT Into a Senior Data Analyst
I once spent an entire weekend building a churn model I was genuinely proud of. Clean features, honest validation, a precision number I wanted to frame.
The VP looked at it for nine seconds and asked: "so what do we do Monday?"
I had nothing. The model had an answer. I didn't.
That failure sent me back to how I actually learned analysis, which was briefing senior analysts and watching them work. They never started with code. They started with questions like "what decision does this change?" and "what would make this wrong?" Years later I realized I could brief ChatGPT the same way. Not "write me SQL." Instead: audit this, attack this, tell me why this is naive.
Most people use ChatGPT like a vending machine. Query in, chart out. Then they're surprised when it hallucinates. The ten prompts below are briefings, and that difference is the engine. Anaconda's State of Data Science surveys have found, year after year, that loading and cleaning data eats close to half a data scientist's time. These prompts get some of that week back.
If you already live in pandas and SQL, this list is for you. If you're still learning what a groupby does, bookmark it for a month from now.
1. The Data Audit
Before you trust a dataset, make someone else look for the bodies.
You are a senior data analyst. Profile this dataset and return: 1. Row/column counts, dtypes, null % per column, cardinality per column 2. Duplicate-row estimate and the key you would dedupe on 3. The 5 data-quality issues most likely to break downstream analysis 4. A "do not use this column for ___" list, with reasons Output as a markdown table plus a short verdict list. Do not fabricate values. Mark anything you cannot verify as UNVERIFIED. tip: Paste 50 to 100 sample rows plus your schema. It cannot audit what it cannot see.
2. The Cleaning Plan
Cleaning decisions are analysis decisions. Never let anyone "just clean it."
Take the issues from the audit above. Produce a cleaning plan as a table: issue | exact transformation | risk of applying it | how to verify it worked | rollback. Apply NOTHING yet. Flag any transformation that changes the meaning of a metric rather than its format. 3. The SQL Pair
One readable query, one fast query. The gap between them is a free lesson about your schema.
Business question: <question> Schema: <tables> Write two SQL queries that answer it: 1. READABLE, optimized for a human reviewing it in a PR 2. FAST, optimized for the engine Explain the plan difference in 3 bullets. State every assumption you made about grain, filters, and nulls. tip: Give it a wrong-but-plausible query you already have and ask why it returns 3x the expected rows. The diagnosis teaches more than the answer.
4. The Exploratory Roadmap
When was the last time an exploration changed a decision? If you can't remember, you're sightseeing, not analyzing.
Given this schema and this business goal: <goal> List the 8 exploratory questions I should answer, in priority order. For each: the plot type, the decision it changes, and how long it should take. Skip any question that wouldn't change a decision if answered. 5. The Statistician
Most analysis errors I've caught in review were test selection errors, not code errors.
Question: <what you want to know> Sample: <n, groups, how collected> Recommend the exact statistical test. List the test's assumptions, how to check each one on my sample, and the effect-size measure you would report. If any assumption fails, give the nonparametric alternative and what I lose by using it. End with the one sentence you would say to a skeptical reviewer. 6. The Chart Critic
Your chart makes sense to you because you made it. This is the reviewer you don't have on staff. When was the last time you showed your chart to someone whose job was to break it?
Here is my chart spec: <chart type, encodings, axes, sample data> Critique it as a data-viz reviewer: what it hides, what could mislead, which encoding choice is doing the most damage, and what a colorblind reader sees. Then give the revised spec. Keep the revision minimal. One chart, one message. 7. The So-What Pass
This one pays for the whole list. Results without consequences are trivia.
Here are my analysis results: <findings> For each finding give: the business consequence, your confidence (high/med/low with reason), what evidence would falsify it, and the strongest objection a skeptic would raise. Rank findings by decision-relevance, not statistical significance. Cut anything that survives only because it is interesting. tip: "Rank by decision-relevance, not significance" is the sentence that makes this prompt senior instead of junior.
8. The Executive Summary
Five bullets, one number each, zero methodology. Harder than it looks. That's the point.
Compress this analysis into a 5-bullet executive summary for a non-technical leadership team. Each bullet: one number, one consequence, one owner-ish verb. No methodology, no hedging language, no "further analysis is needed." The last bullet must state what we do on Monday. 9. The Anomaly Autopsy
When a metric moves, the first hour is wasted if you can't decompose the movement. This prompt is that hour.
Metric <X> dropped <Y>% week over week. Dimensions available: <list> Produce a decomposition plan: how much of the drop is mix shift vs rate change, the three segments most likely to explain it, and the exact query to confirm each. Then tell me which alternative explanation I am probably ignoring. 10. The Flip Test
Run this before you present. It's the hostile reviewer, on demand, an hour before the real one shows up in the meeting.
Attack my conclusion like a hostile reviewer: <conclusion> Go through: selection bias, reverse causality, multiple-comparisons risk, Simpson's paradox, timeframe cherry-picking, survivorship bias. For each: plausible or not for THIS analysis, and the check that would settle it. End with the single objection most likely to be right. The Honest Limitation
These prompts produce briefings, not analysis. The model doesn't see your full data, doesn't know your business until you tell it, and will follow a bad framing as happily as a good one. Every number it hands back is a hypothesis until you've run the query yourself.
The alternative deserves its turn too: sometimes the right move is no prompt at all. Open the notebook, plot the raw data, look at it. When the data is small and the question is fuzzy, your eyes beat any briefing.
What's the one prompt you'd add as #11, the one you actually paste week after week? Drop it in the comments. The best one becomes the follow-up article.
