When to use it
You have an export (a CSV, a spreadsheet, a dump from a CRM or analytics tool) and you need to work out fast "what's actually going on here" — before building a dashboard. The AI plays product analyst. The result: a map of the data, hypotheses, specific findings with numbers, and a list of charts to build — not "things look fine overall".
The prompt (copy and paste)
You are a product analyst. Break down the dataset below.
DATA: <PASTE CSV / a table / attach the file>.
CONTEXT: <what this data is, what period it covers, the business question — e.g. "why did sales drop">.
Work through this in order:
1. DATA MAP: list the columns, the type of each, units, ranges and unique values. Flag missing values and suspicious ones (outliers, zeros, duplicates).
2. QUALITY: what gets in the way of analysis (gaps, inconsistent date formats, anomalies) and how to account for it.
3. FINDINGS: the 3-5 main observations — EACH with a specific number from the data (a share, an average, a rise or fall, the top segment).
4. HYPOTHESES: 2-3 explanations for the findings, plus what extra cut or data you'd need to test them.
5. CHARTS: 3-4 visualisations that show the picture best (chart type + what's on the axes + why).
Rely ONLY on the data provided. If something isn't visible in it, say so rather than filling in the blanks.
Why it works: a strict EDA structure plus the requirement of a number for every finding stops the model sliding into vague phrases like "the trend is positive".
Filled-in example
Context: an export of 5,000 orders over a quarter, question: "why has average order value dropped". Data: a CSV with the columns date, channel, items, total, new_customer.
What the AI should return: the map (date — a date, total — roubles, 200-40000, 3% missing in channel); quality (dates in two formats — normalise to ISO); findings ("the share of orders from the 'discount' channel rose from 12% to 31%", "average order value there is 40% lower", "new customers account for 70% of the cheap orders"); hypothesis "the promo brought in low-value traffic"; charts (a line of weekly average order value; a stacked bar of channel share; a boxplot of order value by channel).
Variations
- Straight to code. "Give me Python (pandas + matplotlib) that computes the findings and plots the charts" — for reproducibility.
- Cohort cut. "Split users into cohorts by month of first purchase and show retention."
- Anomalies only. "Find the top 20 outlier rows and explain what's suspicious about them" — for data cleaning.
Pro tips
- With large files, don't paste everything into the chat — give a 50-100 row sample plus a description of the columns, or attach the file in a data-analysis mode (Code Interpreter / advanced data analysis) so the model computes on real numbers instead of eyeballing.
- Requiring a number for every finding is the main defence against fluff: a sentence with no figure from your data is a cue to ask "based on what?".
- Hypotheses are not conclusions: have the model flag what still needs checking. A correlation inside one CSV is not proof of cause.