Research · Data & analysis
The Best Prompt for Data Analysis
The best prompt for data analysis makes the model check the data before it interprets it and show how each number was produced. This prompt reads the table back, flags quality problems, plans the calculations, then reports findings with the formula, inputs and sample size behind each. A separate section states what cannot be concluded, so you receive an answer to your question together with its limits.
Show-Your-Working Data Analysis Prompt
Analyzes a pasted table or dataset step by step: data check, calculations with working shown, findings tied to those calculations, and what the data cannot support.
- Works with
- ChatGPT, Claude, Gemini
- Difficulty
- Intermediate
- Complexity
- Multi-part reply
- Variables
- 5
- Version
- 1.0
Customize this prompt
Fill in what you know. Anything left blank stays as a [bracketed placeholder]. Nothing you type leaves your browser.
5 variables · fields marked * matter most
# Role You are a data analyst. You show your working, you keep observation separate from interpretation, and you say plainly when the data cannot answer the question. # Inputs Question to answer: [Question to answer] What the data is and how it was collected: [What the data is and how it was collected] Who will read the analysis: [Who will read the analysis] Known issues with the data: [Known issues with the data] <data> [Your data] </data> # Procedure Step 1 - Read back the data. State the number of rows and columns you received, what one row represents, and each column's type and unit. Report missing values, duplicates, inconsistent formats and values that look like errors. Do not correct, drop or fill anything yet; list what you propose to do and why. If the paste looks truncated or misaligned, stop and tell me. Step 2 - Plan. List the calculations that would answer the question and the columns each one uses. If the question cannot be answered with these columns, say so here. Step 3 - Calculate. For every figure you report, show the formula, the values that went into it and the result. Give the count (n) behind every percentage and average. State how you rounded. If you ran code, include it; if you worked it out without running code, say so and mark any result that depends on summing or averaging more than about twenty values as [RECHECK]. Step 4 - Findings. Three to six findings, each one sentence followed by the calculation it rests on. Order them by relevance to the question. Describe size and direction before any explanation of cause. Step 5 - What cannot be concluded. A separate section covering: causes the data cannot establish, groups too small to compare, anything missing that would be needed (a baseline, a control, a longer period), and ways the collection method could bias the result. Step 6 - Next steps. The two or three additional data points or checks that would most reduce the uncertainty. # Rules - Use only the data inside the tags. No outside benchmarks, typical values or industry averages. - Do not claim statistical significance unless you ran a named test and show its inputs and result; with small samples, say the difference may be noise. - Never describe a correlation as an effect. - Outliers: report results with and without them instead of removing them quietly. - Write the findings for the audience named above, but leave the working intact. # Output format Six headed sections in the order above. Use tables for the data read-back and for any grouped results. End with a one-paragraph plain-language answer to the question, including its main caveat.
Use it when you are
- Making sense of a survey export or a small spreadsheet without a dedicated analyst
- Checking an analysis you did by hand before presenting it
- Comparing groups, periods or variants in a results table
- Writing up findings from operational data for a non-technical reader
- Deciding whether a dataset can answer a question before investing more time
Why this prompt works
The data is read back before it is used
Pasted tables arrive with shifted columns, mixed date formats and blank cells. Asking for row and column counts, column types and a list of problems confirms that the model is looking at the same table you are. A mismatch in the row count is the fastest signal that something went wrong in the paste.
Cleaning is proposed, not performed silently
Dropping blanks or correcting an odd value changes the result. The prompt asks the model to list what it intends to do and why, and to report results with and without outliers, so the cleaning decisions stay yours and stay visible in the write-up.
Every figure carries its formula, inputs and n
Language models can make arithmetic slips when calculating in text. Showing the working makes each number checkable in a spreadsheet, the [RECHECK] flag marks the calculations most at risk, and the count behind each percentage stops a result from three rows reading like a trend.
A dedicated section for what cannot be concluded
Left alone, an analysis tends to explain every pattern it finds. A required section on causes the data cannot establish, groups too small to compare and missing baselines gives those caveats a fixed place in the output, where the reader will see them, instead of leaving them to a closing sentence.
No outside benchmarks
A model may compare your numbers with "typical" values it cannot source. Restricting the analysis to the pasted data keeps every statement traceable to a row you can inspect. If you want a benchmark, paste it in as data and say where it came from.
How to use it
- Paste the data as plain text with a header row. Remove names and other personal details you do not need for the question.
- Explain what a row represents and how the data was collected. Collection method drives what can and cannot be concluded.
- Ask one question. Run the prompt again for a second question instead of stacking them.
- Compare the row and column counts in step 1 with your file. If they differ, fix the paste before reading on.
- Approve or change the proposed cleaning, then recheck two or three calculations in a spreadsheet, starting with anything marked [RECHECK].
- Share the findings together with the "cannot be concluded" section, not without it.
AI model compatibility
| Model | Status | Notes |
|---|---|---|
| ChatGPT | Expected to work | For anything beyond a small table, upload the file and use a mode that can run code, then keep the instruction to include the code so you can rerun the calculations yourself. |
| Claude | Expected to work | The prompt places the table in <data> tags, in line with Anthropic's guidance on separating material from instructions. If a code execution tool is available, ask for the calculations to be run with it. |
| Gemini | Expected to work | Paste the table as CSV text with a header row so column boundaries are unambiguous, and keep the six steps in order. |
“Expected to work” means the prompt uses plain, model-independent structure and has been reviewed for that — not that it has been run on that model. A model is marked Tested only when a test record is published below.
Worked example
Scenario: A small online retailer checking whether a courier change affected delivery times, using six months of order data.
# Role You are a data analyst. You show your working, you keep observation separate from interpretation, and you say plainly when the data cannot answer the question. # Inputs Question to answer: Did average delivery time change after we switched couriers on March 1? What the data is and how it was collected: One row per order, January to June. Columns: order ID, dispatch date, delivery date, courier, region. Exported from our order system. Who will read the analysis: Operations director, comfortable with averages but not statistics Known issues with the data: Some February orders have no delivery date because scans failed <data> [About 180 rows of CSV pasted here with a header row: order_id, dispatch_date, delivery_date, courier, region.] </data> # Procedure Step 1 - Read back the data. State the number of rows and columns you received, what one row represents, and each column's type and unit. Report missing values, duplicates, inconsistent formats and values that look like errors. Do not correct, drop or fill anything yet; list what you propose to do and why. If the paste looks truncated or misaligned, stop and tell me. Step 2 - Plan. List the calculations that would answer the question and the columns each one uses. If the question cannot be answered with these columns, say so here. Step 3 - Calculate. For every figure you report, show the formula, the values that went into it and the result. Give the count (n) behind every percentage and average. State how you rounded. If you ran code, include it; if you worked it out without running code, say so and mark any result that depends on summing or averaging more than about twenty values as [RECHECK]. Step 4 - Findings. Three to six findings, each one sentence followed by the calculation it rests on. Order them by relevance to the question. Describe size and direction before any explanation of cause. Step 5 - What cannot be concluded. A separate section covering: causes the data cannot establish, groups too small to compare, anything missing that would be needed (a baseline, a control, a longer period), and ways the collection method could bias the result. Step 6 - Next steps. The two or three additional data points or checks that would most reduce the uncertainty. # Rules - Use only the data inside the tags. No outside benchmarks, typical values or industry averages. - Do not claim statistical significance unless you ran a named test and show its inputs and result; with small samples, say the difference may be noise. - Never describe a correlation as an effect. - Outliers: report results with and without them instead of removing them quietly. - Write the findings for the audience named above, but leave the working intact. # Output format Six headed sections in the order above. Use tables for the data read-back and for any grouped results. End with a one-paragraph plain-language answer to the question, including its main caveat.
What a good response should contain
- A read-back giving the row count, the columns and how many orders lack a delivery date, with a proposal for handling them instead of a silent deletion
- Average and median delivery days for each courier, each with its formula, the number of orders behind it and a [RECHECK] flag if no code was run
- A regional breakdown only where each group has enough orders to compare, with small groups named as too small
- A "cannot be concluded" section noting that the two periods cover different months, so seasonal demand or weather could explain part of any difference
- A closing paragraph that answers the question in plain language with its main caveat attached
This is an illustrative scenario showing how to fill in the prompt and what to look for. It is not a recorded model output. Recorded outputs will be published here once this prompt has a test record.
Testing and editorial review
Status: Untested. Reviewed for structure and clarity by the editorial team. No recorded model test yet.
No test record has been published for this prompt yet. When it is tested, this section will show the model and version, the date, the inputs, the actual output and notes against our seven-point rubric. We do not publish numerical scores.
| Author | TheBestPromptFor Editorial Team — The editorial team writes, reviews and maintains the prompts and guides on TheBestPromptFor. Prompts are drafted with AI assistance and reviewed against published editorial standards. |
|---|---|
| Published | November 15, 2026 |
| Last reviewed | November 15, 2026 |
| Last tested | Not yet tested |
| Version history |
|
Limitations and things to check
- Without code execution, arithmetic over many rows can be wrong. Recheck totals and averages in a spreadsheet.
- Large datasets will not fit in a chat message. Upload the file to a tool that can run code, or analyze a clearly described sample.
- The model only knows what you tell it about how the data was collected, and hidden collection problems will not show up in the table.
- This prompt covers descriptive analysis and simple comparisons. Modeling, forecasting and formal hypothesis testing need a statistician or proper tooling.
- Do not paste personal or confidential data into a tool your organization has not approved for it.
Alternative prompts
Research
Research Paper and Report Summary Prompt
Summarizes a paper or report for a named reader: bottom line, method, key findings with exact figures, limits of the evidence, and what it means for them.
- ChatGPT
- Claude
- Gemini
E-commerce
Customer Review Analysis Prompt
Turns pasted reviews into counted themes with verbatim quotes, causes of complaints, customer language and recommendations graded by support.
- ChatGPT
- Claude
- Gemini
Coding
Python Script and Module Prompt
Turns a plain-language spec into a complete, typed Python script or module, with assumptions, run commands and edge cases listed separately.
- ChatGPT
- Claude
- Gemini
- Cursor
Productivity
Structured Decision-Making Prompt
A two-round prompt: the model questions you first, then weighs your options against your criteria and argues against the one you already favor.
- ChatGPT
- Claude
- Gemini
Related resources
Guide
How to Improve AI Responses
A diagnostic approach to weak AI answers: name the fault, trace it to the gap in your prompt, then follow up precisely or rewrite and start again.
Technical guide
How to Test AI Prompts
A seven-criterion rubric and a nine-field test record for evaluating prompts, plus why this site describes results in words and publishes no scores.
Technical guide
Prompt Chaining: Breaking a Task Into Linked Prompts
How to split a large task into a sequence of focused prompts, design the handoffs between them, and debug the chain when a step goes wrong.
Frequently asked questions
Can an AI model do the calculations accurately?
It depends on how the calculation is done. A model that runs code on an uploaded file computes results the way a script would. A model working from pasted text may slip on long sums. That is why the prompt asks which method was used and flags hand-worked results for rechecking.
Why does the prompt refuse to call a difference significant?
Because the word has a specific statistical meaning. A model may use it loosely for any visible difference. The prompt allows the claim only when a named test was run and its inputs are shown; otherwise it asks for the plain statement that a small-sample difference may be noise.
What format should I paste the data in?
CSV or tab-separated text with a single header row is the most reliable. Avoid merged cells, subtotal rows and multi-line headers, since they blur column boundaries. If you copy from a spreadsheet, check the first few rows of the read-back to confirm values landed in the right columns.
Sources and references
- Prompt engineering overview — Anthropic
- Prompt engineering guide — OpenAI
Was this prompt useful?
No account needed. Do not include personal information.