My three-prompt habit for checking a small CSV

When I want a quick answer from a spreadsheet, I try to keep the first pass to three prompts: inspect the data, calculate one useful comparison, and check the result. That limit makes me choose a question before spending money on a long conversation. Here is a walkthrough using a small export of orders from a student side project.

Choose one question your export can answer

Start with a question that leads to a decision. For this example, ask which product brought in the most revenue last month. Assume the export contains an order date, order identifier, product name, quantity, unit price, and payment status.

Write down what revenue means before opening the chat. Here, use quantity multiplied by unit price for paid orders, excluding canceled orders. If refunds, taxes, or discounts appear separately, decide how to handle them. A precise definition prevents an apparently helpful answer from quietly using the wrong calculation.

Remove customer names, email addresses, and other fields the question does not need. Keep the original export unchanged so you can compare it with any cleaned version later.

Make sure your chosen tool can calculate

Use an application that can read the CSV and execute calculations, or a coding assistant that can run an analysis with your permission. Model access alone does not provide file handling or a calculation environment. If your application only exchanges text, ask it to help construct spreadsheet formulas, then run those formulas yourself.

MyRelay can provide the model connection for a compatible client through one personal key and a shared pay-as-you-go budget. Confirm that your client supports the file and analysis features you need. You can create a MyRelay account to keep the tokens and cost of each request visible while you experiment.

Ask for a data inspection before any ranking

Make your first prompt specific: “Inspect this orders CSV before analyzing revenue. Report the row count, column names, date range, missing values in required fields, and possible duplicates. Identify how payment status is represented. Explain any uncertainty about whether each row represents an order or an order item. Do not calculate product rankings yet.”

Read the response against the actual file. Repeated order identifiers may be normal if one order contains several products. Dates may use an unexpected format, and prices may be stored as text. Resolve these issues before continuing, because they can change the result more than the choice of model.

Request one calculation with explicit rules

Once the structure is clear, send the second prompt: “For the most recent complete calendar month in the file, calculate revenue by product using quantity multiplied by unit price. Include only paid orders and exclude canceled orders. List the five highest-revenue products, their revenue, and their share of total included revenue. State the date boundaries, excluded row count, and total revenue used.”

Adapt the wording to the actual status values and your treatment of refunds. Ask the tool to execute the calculation and distinguish executed results from suggested methods. If it cannot execute anything, move the calculation into your spreadsheet rather than accepting a numerical guess.

Keep the requested output small. Five rows and a short explanation are usually enough to decide whether a deeper analysis would help.

Check the leading result independently

Use the third prompt to make the result easier to audit: “Show the calculation behind the leading product, including its included row count, total quantity, and revenue. Report missing prices, negative quantities, or other exceptions that could affect this result. Explain whether the totals reconcile.”

Then filter that product in your spreadsheet, apply the same dates and payment rules, and calculate its revenue yourself. Also compare the overall included revenue with a separate spreadsheet total. Agreement provides useful evidence; asking the model to repeat its answer provides much less.

Save the method and watch the request costs

Record the question, date boundaries, filtering rules, verified totals, and any unresolved exceptions. Next month, reuse those instructions with the new export. Avoid resending unrelated history or asking for a lengthy report when a small table answers the question.

In MyRelay, review the tokens and cost of the three requests. Use that record to set a realistic allowance for future checks. Spend another request when a discrepancy or a new decision justifies it, and keep the verified calculation alongside the written explanation.