Data Analysis Prompts for Questions You Can Verify

Good data analysis prompts state the question, define the data, and require a reproducible calculation with explicit checks. Ask the assistant to inspect the dataset before interpreting it, and distinguish work it actually executed from code it merely suggested. A clear result includes its exclusions and uncertainty, not just a confident conclusion.

Tools like fenbs keep reusable analysis prompts alongside project assumptions, so a fresh conversation does not silently redefine the same metric. Keep the actual dataset in its approved environment and bring only the context needed for the task.

What belongs in the analysis instructions?

Begin with a decision the data might inform. “Compare completed service visits by region” is narrower than “find insights.” Define what each row represents, when the data was collected, and whether the file is complete. A sample can help test a method, but it cannot establish totals for the full population.

Give the assistant a data dictionary: a short explanation of each column, its units, and allowed values. Name your chosen spreadsheet, query language, or analysis environment. If it cannot run calculations there, ask for a method and label the numerical result unverified until you execute it.

What prompt should you use to inspect a dataset?

Inspection should happen before cleanup. Otherwise an assistant may remove unusual rows that turn out to be the most relevant part of the question. The NIST explanation of exploratory data analysis↗ (opens in a new tab) includes discovering structure, finding anomalies, and examining assumptions. Use that as a reason to look first and explain transformations later.

Inspect [dataset or approved sample] before answering [question].
Each row represents [unit]. Data dictionary: [definitions].
Report row and column counts, detected types, date coverage, missing values, duplicate candidate keys, unexpected categories, and unusual numeric values.
Do not delete, fill, or convert anything yet. Separate observed facts from suspected problems. For each proposed cleaning action, explain how it could change the answer. Show executable inspection steps and say which ones you actually ran. If you cannot access the file, stop and request the needed information.

A missing-value count needs a definition too. For example, the pandas documentation for missing-value detection↗ (opens in a new tab) says empty strings are not automatically treated as missing by that method. Ask whether blanks, spaces, “unknown,” and other placeholders occur, rather than assuming one function captures every kind of absent information.

Save agreed column meanings in the fenbs project context and Glossary, so future analyses begin with the same definitions.

How do you ask for a calculation you can reproduce?

Specify the numerator, denominator, grouping, and exclusions in ordinary language. “Average revenue per customer” can mean an average of customer totals or an average of transaction values; those answer different questions. Ask the assistant to restate the formula before applying it, especially when joining multiple tables.

Answer [specific question] using these approved definitions: [definitions].
Numerator: [definition]. Denominator: [definition].
Group by: [columns]. Include: [rules]. Exclude: [rules].
Before calculating, restate the formula, row unit, time boundary, and join keys. Flag unresolved ambiguity.
Then show the executable calculation, intermediate counts, final values with units, and the handling of zero denominators and missing values. Keep full precision until display. Distinguish executed output from an expected result. Do not infer causes from a difference between groups.

In a small invented check, Region A has eight completed visits out of ten scheduled visits, while Region B has six out of twelve. The completion shares are 80 percent and 50 percent. Their combined share is fourteen out of twenty-two, about 63.6 percent; averaging the two regional percentages would answer a different question.

Use a tiny example like this to check the method, then run it on the approved data. If a join expands ten visits into twenty rows, inspect the relationship before trusting totals. A correct formula applied to duplicated records still gives the wrong answer.

What should a chart prompt include?

Tell the assistant what the reader needs to compare. A trend over time, a distribution of values, and a comparison of categories need different displays. Ask it to explain the proposed chart and provide the table behind it, so the visual remains connected to numbers you can inspect.

Propose a chart to answer [question] for [audience].
Use the validated result table below: [table].
Explain why the chart suits the comparison. Specify title, axis labels, units, sorting, time interval, and missing-data treatment. Show group sample sizes where relevant. Do not combine incompatible units or hide an exclusion. Return the underlying plotted values and a plain-language description. If the data cannot support the requested comparison, explain the limitation instead of making the chart.

For the service example, a bar chart of completion shares could show the two regions, with scheduled-visit counts in labels or nearby text. A line chart would suggest a sequence that the region labels do not have. Neither display can explain why one region differs without additional evidence.

Read the title as carefully as the picture. “Regional completion shares” describes the measure. “Region B staff underperformed” adds a conclusion about people that this dataset has not established. Keep the visual’s language inside the evidence.

How do you catch a convincing wrong answer?

Require a validation pass that uses explicit checks rather than another request to “be accurate.” Compare totals with the original source, inspect a few individual records, and test an edge case. An assistant can produce a fluent explanation of an incorrect calculation, so fluency is not a check.

• Reconcile row counts before and after each filter or join, explaining every difference.
• Recalculate one group independently with a calculator, spreadsheet, or separate query.
• Test a blank, a duplicate, a zero denominator, and a value on the date boundary.
• Confirm that percentages use the intended population and that displayed units match the source.
• List what would change the conclusion, including incomplete coverage or an unresolved definition.

Record the question that remains after those checks. A difference may be real in this file while still being too uncertain to guide a wider decision. Our AI data analysis review guide covers the broader workflow; this page provides prompts for its individual steps.

Keep a reusable validation checklist with your fenbs Skills and prompts. Update it when a reviewed analysis reveals a missing check.

What should you save for the next analysis?

Save the dataset identifier, extraction date, approved definitions, prompt version, and outstanding checks. Record the path or reference appropriate to your environment without copying restricted records into a general work note. A future assistant needs to know what was analyzed and which decisions shaped the calculation.

Separate a reusable method from one run’s results. The prompt can say how to calculate a completion share; the run record says which file and time period it used. System prompt examples explain reusable instructions, while project-management prompts help turn an unresolved check into follow-up work.

How to do this with fenbs

1. Create a free account and add the analysis project, with a concise description of the question.
2. Save the reviewed prompt templates in Skills and prompts. Include required inputs and what counts as a checked result.
3. Record approved dataset assumptions and metric definitions as project notes or Decisions and rules, according to whether they are facts or choices.
4. Connect your assistant through Integrations. Ask it to read the context and chosen prompt, then work in your approved analysis environment.
5. Save progress with the dataset reference, completed checks, unresolved issues, and next step. Keep the data and executable analysis in the systems intended for them.

fenbs does not execute statistical analysis or replace a spreadsheet, database, or notebook. If you only need to calculate a single number, those tools may be enough. Its AI brain retains the shared work instructions; the AI prompts hub has other reusable patterns.