Data-to-Insight Analyst

What it does

You paste a small dataset (an export, a report table, a few columns from a spreadsheet) and one line on the decision you face. Instead of describing your data back to you, it answers the decision first, for the period the decision covers, and then shows the numbers behind the answer: one compact table with every option you are choosing between, and the findings that carry the answer.

Before it draws a conclusion it runs the checks an experienced analyst would:

• partial periods, duplicates, blanks and impossible values;
• outliers, with each conclusion recomputed without them;
• comparisons that are not like for like;
• whether a gap between two averages is bigger than the uncertainty of those averages, and how much more data would settle it if not;
• before/after comparisons that cannot separate a change from everything else that moved.

It ends with a first step sized to your decision. The success threshold comes with its derivation (break-even, the previous level or a margin you gave), so it is not a number picked out of the air.

Use it when

• You have a weekly or monthly export (sales, traffic, bookings, support tickets, costs) and a real decision to make from it.
• You want to know which differences in your numbers are real and which are normal ups and downs.
• You suspect the data itself has problems: a missing month, a duplicated row, one strange day.
• You changed something (a price, a process, a supplier) and want to know whether to keep the change.

Skip it when

• The dataset has thousands of rows. Use a spreadsheet, a notebook or a code-running analysis tool, and bring the summary here.
• You need a formal statistical analysis (models, forecasts with intervals, results you will publish). This gives a practical read with rough uncertainty, not a formal analysis.
• You have no decision in view and only want a description of the data. A plain request does that well.

The prompt

DATA:
[PASTE DATA]

CONTEXT: [CONTEXT]

Analyse the data above to help me make the decision in the context line, for the period that decision covers. If the context names no clear decision, infer the most likely one, state it in one line and answer it. Do not ask me questions first: where something is missing, make the most reasonable assumption, say so in a few words where you use it, and finish.

Before you write, check the data the way a careful analyst would, and use what you find:
- Options: identify every option the decision chooses between (each product, channel, period, group or setting in play) and compare all of them, not only the two biggest.
- Data problems: duplicate rows, blanks, impossible values, unit or currency changes, and partial periods (a month cut short, a week with missing days). Say how you handled each one.
- Outliers: find them, leave each one out and recompute, and say whether the conclusion still holds. Do not let one unusual row drive an average.
- Fair comparisons: compare like with like (same weekdays, same number of days, same season). Use per-day or per-unit figures when periods differ in length.
- Real differences: judge a gap between two averages or ratios against the uncertainty of those averages, never against the spread of single rows. That uncertainty is roughly the spread of the single values divided by the square root of their count, and for a gap between two averages it combines as √(u1² + u2²). A gap under about twice that is not settled: say so, do not rest the recommendation on it, and estimate how much more data would settle it.
- Cause: when two measures move together (for example overtime hours and error counts), say that they move together, not that one causes the other. A before/after comparison mixes the effect of a change with everything else that changed over the same period.

Then write the answer, usually 300-500 words (shorter when the answer is simple, longer only when the decision has many options), with no filler and no restating of my context, in this order:
1. Answer: one or two sentences that answer the decision directly, for the period it covers. "Not enough evidence yet" is a valid answer when it is the honest one; then say what would settle it.
2. What the data shows: one compact table with the key metric for each option the decision chooses between (summarise; do not echo the raw rows), then the findings the answer rests on, most decision-relevant first, as many as the data supports. Show the arithmetic for every number the answer depends on (a ÷ b = c), and check every range and total against the rows.
3. What looks wrong or unusual: the data problems and outliers you found, how you handled each, whether the answer changes without it, and what to check in the source.
4. First step: the first concrete step toward the decision, sized to it: the metric to watch, the threshold that counts as success and how you derived it (from the data, a break-even point, the previous level or a margin I gave), and when to look again. If a figure the threshold needs is missing, name it and give the formula instead of a number. If the step is meant to show whether a change works, use a comparison that separates its effect from noise (a held-out or matched group, or alternating periods), not a before/after check.
5. Limits: what this data cannot tell me, and the one extra piece of data that would most change the answer.

Close with at most three lines on anything else I need to act on the answer that the sections above do not cover; leave this out when there is nothing.

How to use

1. Replace the two placeholders:
   • [PASTE DATA]: your CSV text or table, with its header row. Tab-separated text copied straight from a spreadsheet works.
   • [CONTEXT]: two or three lines on what the business is, what one row means, the decision you face and the period it covers. Add any cost or margin you know (what a lost sale, a missed slot or an hour of staff time is worth): the first step's threshold is derived from it. For example: "Physiotherapy practice, one row per week. Deciding whether to keep SMS reminders for the rest of the year."
2. Paste the whole thing as the first message of a new chat.
3. Read the Answer line first. Before you act, check one or two of the cited numbers against your sheet.
4. Every assumption is marked where it is used. Check those before you act on them.
5. Want it shorter? Change "300-500 words" in the prompt. A tighter range keeps the answer on the decision but leaves less room for the comparison table.

How to run (Oct 2026)

• ChatGPT: paste into a new chat on GPT-6. The Medium position of the Thinking slider is enough for a few dozen rows; move it to High when there are many columns or the decision is close. Typing "think harder" does not change the level. To reuse the prompt, put it (without the data) in a ChatGPT Project's instructions and paste only the data in each chat.
• Claude: paste into a new chat. Opus 5.5 at its default effort handles this; Sonnet 5.5 is the faster option. The data sits above the instructions on purpose: Anthropic recommends putting long inputs first and the request last. To reuse it, keep the prompt in a Project's instructions.
• Gemini: paste into the Gemini app or Google AI Studio. Google also recommends context first and the question last, which is how the prompt is laid out. To reuse it, save it as a Gemini app skill.
• DeepSeek, Qwen, GLM: send it as the first message through the API or the app you use. Thinking is on by default on DeepSeek-V4.1-Flash and Qwen3.8 and always on for GLM-5.3; keep it on for this job. GLM-5.3 reads text only, so paste the data as text, not as a screenshot.

Example

Input

• [CONTEXT]: Physiotherapy practice, one row per week. From week 7 we sent appointment reminders by SMS (€0.06 each) instead of phone calls (about 3 receptionist hours a week at €22 an hour). A missed appointment costs us a €50 slot. Deciding whether to keep SMS reminders for the rest of the year.
• [PASTE DATA]:

week,reminder,working_days,appointments,no_shows
1,phone,5,118,7
2,phone,5,122,13
3,phone,4,96,9
4,phone,5,121,11
5,phone,5,119,22
6,phone,5,124,8
7,sms,5,126,10
8,sms,5,123,5
9,sms,5,127,9
9,sms,5,127,9
10,sms,5,125,6
11,sms,5,124,11
12,sms,5,122,7

Example output (illustrative, written by Meevra to show the target format; your model's wording will differ)

Answer: Keep SMS reminders for the rest of the year. They cost about €58 a week less than phone calls and would only lose money if they caused more than about 1.2 extra no-shows a week; the data points the other way. Whether SMS itself lowers no-shows is not settled yet, and the decision does not depend on it.

What the data shows

1. SMS saves €66.00 − €7.50 = €58.50 a week. To cancel that out it would have to add €58.50 ÷ €50 = 1.17 no-shows a week, which is 1.17 ÷ 125 = 0.94 points of no-show rate.
2. The gap is 8.26 − 6.43 = 1.83 points in favour of SMS, but it is not settled. Weekly rates swing by about 2 points around each average, so the phone average is uncertain by about 2.0 ÷ √5 = 0.9 points and the SMS average by 1.9 ÷ √6 = 0.8 points; combined, √(0.9² + 0.8²) = 1.2 points. The gap is 1.83 ÷ 1.2 = 1.5 times that, under the rough bar of 2. About 1.7 times as many weeks of each method (9-10 weeks each) would settle it.
3. Even the pessimistic end of that range, SMS 1.83 − 2 × 1.2 = 0.6 points worse than phone, stays below the 0.94-point break-even.

What looks wrong or unusual

• Week 9 appears twice with identical values. Counted once.
• Week 3 had 4 working days. Rates per appointment keep it comparable, so it stays in.
• Week 5: 22 ÷ 119 = 18.5% no-shows, about five times the usual weekly swing above the phone average. Left out of the table; with it, phone is 70 ÷ 700 = 10.0% and the gap 3.6 points. The answer is the same either way; only the size of any SMS effect changes. Check the diary for that week (a closure or a weather day would explain it).

First step Keep SMS and look again in 8 weeks. Metric: the no-show rate over those weeks. Success: at or below 9.2%, which is the phone level of 8.26% plus the 0.94-point break-even; above that, SMS costs more than it saves. If you later want to know whether SMS itself lowers no-shows (for example before paying for a second reminder), send the extra reminder to a random half of patients over the same weeks and compare the two halves, not this year against the past.

Limits Six weeks per method, and SMS ran later in the season than phone calls, so seasonal changes are mixed into the comparison. The extra data that would most change the answer: no-shows for the same weeks last year.

Troubleshooting

• It compared only some of the options. Reply "Put every option the decision chooses between in the table and say whether the answer changes."
• It calls a small gap real, or dismisses a real one. Ask "Compare the gap with the uncertainty of the two averages (spread of single values ÷ √count), not with single rows, and say whether it is settled."
• The success threshold looks arbitrary. Reply "Derive the threshold from break-even, the previous level or the data, and show the calculation."
• It answered for a shorter period than your decision. Reply "Answer the decision for the full period it covers, then give the first step."
• A number doesn't match your sheet. Ask "Show the two numbers behind each ratio you cited", then correct the misread row and ask for a rerun.
• It asks you questions before answering (common on GPT-6). Reply "Use your best reading, state it in one line, and answer."

© 2026 Meevra · meevra.com · Licensed to the purchaser for personal and internal business use.