ChatGPT Prompts for Data Analysis: 15 Templates for CSV, Excel and SQL (2026)

I have uploaded [filename]. Before any analysis, profile it for me: row count, column names with inferred types, the share of missing values per column, cardinality of each categorical column, min/max/median for each numeric column, and the date range of any date column. Then list the five things about this file that would most likely break an analysis, and tell me which columns you do not trust and why. Show the code you ran.

This file is an export from [system, e.g. Stripe, HubSpot, Shopify]. Clean it for analysis: standardize the date columns to ISO dates, coerce numeric columns stored as text, strip currency symbols and thousands separators, normalize inconsistent category labels (show me the mapping you chose before applying it), and flag exact and near-duplicate rows rather than dropping them silently. Output the cleaned file plus a short changelog of every transformation and how many rows each one touched.

Write a [Excel / Google Sheets] formula that does this: [describe the result you want in plain language]. My data is in [sheet and range], with these columns: [list columns and what they contain]. Give me the formula, a one-line explanation of each part, and the two ways it will break (blank cells, text stored as numbers, ranges that grow). If a helper column would make it simpler and more auditable, say so and show that version too.

I want to answer this question with a pivot: [question]. My columns are [list]. Specify the pivot: what goes in rows, columns, values, and filters, which aggregation to use for each value field, and how to handle blanks. Then tell me what this pivot will hide, and which second pivot I should build to check the first one is not misleading me.

Here are the results of an A/B test: [variant names, users per variant, conversions per variant, test duration]. The metric I care about is [metric]. Calculate the observed lift, the confidence interval, and the p-value, and state plainly whether this is a result I should act on. Then list the reasons this test could be wrong regardless of the statistics: sample ratio mismatch, novelty effect, insufficient duration, peeking, segment mix shift. Show your working.

Build a monthly cohort retention table from this data: [describe the file or paste it, noting the user id, signup date, and activity date columns]. Show retention as both counts and percentages, cohort by cohort. Then interpret it: which cohorts are healthier, where the drop-off concentrates, and whether the trend across cohorts is improving or deteriorating. Name the one thing in this table you would investigate first.

I want to communicate this finding: [state the finding in one sentence]. The underlying data is [describe or upload]. Recommend the chart type and explain why it beats the two obvious alternatives, then build it with clear axis labels, no chart junk, and the finding readable without a caption. If the honest chart is less impressive than the flattering one, build the honest one and tell me what the flattering version would have hidden.

Here is a time series of [metric] by [day/week]: [upload or paste]. Identify the anomalies, separating real level shifts from noise and from seasonality. For each anomaly, give the date, the size relative to the baseline, and three candidate explanations ranked by how likely they are, including the boring ones like a tracking change or a data pipeline failure. Tell me what other data I should pull to distinguish between them.

Write Python (pandas) that does this analysis: [describe the steps and the output you want]. My dataframe has these columns and dtypes: [paste]. Write it as readable, commented code that a colleague could modify, avoid chained assignment, and after the code explain each transformation in one line and state every assumption it makes about the data. Then show me the output on the first few rows so I can check it before running the whole thing.

I ran this analysis and reached this conclusion: [describe the method, the data, and the conclusion]. Review it as a skeptical senior analyst. Where is the reasoning weakest, what confounders did I ignore, is the sample adequate for the claim, and which step would a reviewer attack first? Give me a list of specific checks I can run to either strengthen or kill the conclusion, in the order you would run them.

Turn this analysis into a one-page readout for [audience, e.g. the exec team]: [paste findings and numbers]. Structure it as the headline finding in one sentence, three supporting points with the numbers attached, the main caveat stated honestly, and a recommended decision with what would change your mind. No jargon, no chart descriptions, nothing that needs the appendix to make sense.

I want to know whether [describe the comparison, e.g. average order value differs between three regions]. The data: [describe the variables, their types, the sample size per group, and whether the same people appear in more than one group]. Recommend the statistical test, explain in two sentences why it fits and why the obvious alternative does not, list the assumptions it needs and how to check each one on my data, then run it and report the result with an effect size, not just a p-value. If the sample is too small for any test to be meaningful, say so.

Fit a [linear / logistic] regression of [outcome] on [predictors] using the uploaded file. Before fitting, check for missing values, obvious outliers, and predictors that are near-duplicates of each other. Report each coefficient in plain language (holding the others fixed, a one-unit change in X goes with ...), its confidence interval, and how much of the variation the model explains. Then list what this model cannot tell me: causation, effects outside the observed range, and any variable I left out that could be doing the work.

Here is [metric] by [week/month] for the last [n] periods: [upload or paste]. Forecast the next [n] periods using a simple, explainable method (seasonal naive or exponential smoothing before anything fancier). Show the forecast with an 80% range, the method in one paragraph, and a backtest: what the same method would have predicted for the last [3] periods against what actually happened. Flag any period in the history that looks like a one-off and say whether you excluded it.

[Metric] moved from [old value] to [new value] between [period A] and [period B]. Using the uploaded data, decompose the change: how much comes from volume versus rate, and how much from each [segment, e.g. channel, region, plan]. Separate mix effects (the segments changed size) from rate effects (the segments changed behaviour). Present it as a table that adds up to the total change, then name the one or two components that explain most of it and what I should check to confirm each.