Analyze [DATASET DESCRIPTION] to answer [BUSINESS QUESTION] for [DECISION MAKER]. The available fields, tables, data types, and grain are [SCHEMA AND GRAIN]. The date range, population, and known exclusions are [SCOPE AND EXCLUSIONS]. The metric definitions or business rules are [METRIC DEFINITIONS]. I will work in [ANALYSIS TOOL]. First, restate the unit of analysis and distinguish descriptive questions from causal claims. Identify the minimum data needed and any ambiguity in the business question. Produce: 1. An analysis plan with the population, time window, primary metric, comparison group or baseline, dimensions to segment by, and a short rationale.
2. A data-quality checklist covering duplicate records, missing values, invalid dates, inconsistent categories, join cardinality, timezone or fiscal-calendar issues, outliers, and denominator changes. 3. The exact SQL, Python, spreadsheet formulas, or tool-specific steps needed to produce the analysis. Use readable names, explicit filters, and comments. If a join is needed, state its expected cardinality and how to test for row multiplication. 4. A results template with a metric table, recommended chart type, and a plain-English narrative that separates facts, interpretation, and next questions. 5. Sensitivity checks: at least one alternate date window or threshold, and a check for a segment that could reverse the aggregate result. 6. A short list of limitations and what additional data would be needed before making a causal, financial, or policy decision. Ask up to 3 clarifying questions only if a required input is missing. Before answering, check that each formula’s numerator and denominator use the same population and that every conclusion is supported by an output the proposed analysis can actually produce. Do not invent results, column values, significance, or causal explanations.
Fill in
| Placeholder | What to enter | Example |
|---|---|---|
| [DATASET DESCRIPTION] | Describe what the dataset records and where it came from. | Shopify orders export joined to email campaign sends for an online stationery store |
| [BUSINESS QUESTION] | State the specific decision or question the analysis should answer. | Which email campaign types are associated with the highest 30-day repeat-purchase rate? |
| [DECISION MAKER] | Name the role or team that will use the analysis. | The lifecycle marketing manager deciding next quarter’s campaign calendar |
| [SCHEMA AND GRAIN] | List tables or fields, their data types when known, and what one row represents. | orders: one row per order with customer_id, order_date, net_sales; sends: one row per recipient-send with customer_id, campaign_id, send_time, campaign_type |
| [SCOPE AND EXCLUSIONS] | State the time period, population, geography, and records to exclude. | US orders from January through June 2026; exclude test orders, refunds, and employees |
| [METRIC DEFINITIONS] | Define key metrics, business rules, and known edge cases. | Repeat purchase means a later completed order by the same customer within 30 days of a send; a customer can receive multiple campaigns |
| [ANALYSIS TOOL] | Name the tool where you need formulas, code, or queries. | BigQuery SQL |
How to use
- Write down what one row means before pasting the schema; this prevents accidental double counting.
- Paste the prompt with field names exactly as they appear in your source, then run the data-quality queries before the final analysis.
- Inspect a few raw records behind every surprising segment or chart to verify joins and dates.
- Follow up with: “Write a validation query that compares this result with a customer-level, de-duplicated calculation.”
Variations
SQL debugging
Use this when a query returns the wrong total, duplicates, or an error.
Debug this [SQL DIALECT] query: [QUERY]. The expected grain is [EXPECTED GRAIN], the actual symptom is [SYMPTOM], and the relevant schemas are [SCHEMAS]. Explain the likely failure mode before proposing a corrected query. Check joins for one-to-many multiplication, filters applied after a left join, null behavior, date boundaries, aggregation level, and distinct-count misuse. Return a corrected query with comments plus two small diagnostic queries that prove row counts and totals at each stage. Do not fabricate table contents. Ask up to 3 questions only if missing schema details prevent a safe fix.
Executive summary
Use this when analysis is complete and you need a decision-ready readout.
Turn these analysis outputs into a one-page decision memo for [AUDIENCE]: [RESULTS AND TABLES]. The decision is [DECISION]. Write a headline, 3 to 5 evidence-backed findings with exact numbers, implications, recommended action, risks, and next measurement date. Separate observed results from hypotheses. Include a small table only if it makes the decision clearer. Flag missing denominators, small samples, or comparisons that cannot support the recommendation. Do not add facts or explanations not present in my evidence.
Experiment analysis
Use this when assessing an A/B test or controlled experiment.
Create an A/B test analysis plan for [EXPERIMENT]. Variant A is [CONTROL], Variant B is [TREATMENT], the primary metric is [PRIMARY METRIC], and guardrails are [GUARDRAILS]. Use [DATA SOURCE AND TOOL]. Define the exposure population, assignment unit, attribution window, exclusions, sample-ratio mismatch check, and how to handle repeat exposure. Provide reproducible calculations or queries, a result table template with counts and rates, and guidance on uncertainty. Do not claim statistical significance without the underlying data and method.
Tips
- Declare dataset grain before choosing an aggregate; one row per order and one row per customer require different denominators.
- Test joins with row counts and distinct IDs before calculating revenue or conversion metrics.
- Use a fixed metric definition for every segment; changing eligibility across groups can create a false comparison.
- Treat a correlation as a lead for further investigation unless the analysis has a credible causal design.
FAQ
Can this prompt analyze a CSV directly?
Yes, if you upload it to the tool and provide its columns and row meaning. The prompt asks for checks before drawing conclusions.
Should I ask for a chart first?
Define the decision and metric first. The appropriate chart follows from whether you need a trend, comparison, distribution, or relationship.
Will AI know if my results are statistically significant?
It can calculate or explain a suitable method when you provide the data and assumptions. It should not claim significance from a summary without the necessary counts and design details.