Build a copy-paste AI prompt pack for a data analyst working on [ANALYTICS CONTEXT]. Our typical data sources and tools are [DATA SOURCES AND TOOLS], our key metrics are [KEY METRICS], and the audience for findings is [STAKEHOLDER AUDIENCE]. Follow these rules: [DATA GOVERNANCE RULES]. Produce 15 complete prompts, grouped under these labels: data profiling, SQL and transformation, metric definition, exploratory analysis, anomaly investigation, experiment analysis, dashboard QA, stakeholder communication, and analyst workflow. Include at least one prompt for each label. Each prompt must have: - A short task title. - The exact copy-paste prompt in first person, with 2-5 necessary [PLACEHOLDERS].
- A one-sentence note on what source material to paste. - A one-sentence validation step the analyst must perform outside the model. Make prompts operational. SQL prompts must request the dialect, table grain, join keys, date boundaries, null handling, and a reconciliation query. Metric prompts must separate numerator, denominator, eligibility, time window, and known exclusions. Analysis prompts must require the model to distinguish observation from causal claim and name plausible confounders. Dashboard prompts must check filters, time zones, freshness, duplicated records, and totals against a trusted query. Do not request raw customer data, credentials, or unapproved confidential fields. Do not fabricate query results, benchmarks, or statistical significance. Before answering, self-check that each prompt specifies a concrete deliverable and includes a validation action. Ask up to 3 clarifying questions only if a required input is missing.
Fill in
| Placeholder | What to enter | Example |
|---|---|---|
| [ANALYTICS CONTEXT] | Describe the business, product, and types of questions the analyst supports. | Subscription meal-planning app; growth, retention, and subscription revenue analysis |
| [DATA SOURCES AND TOOLS] | List approved warehouses, BI tools, files, SQL dialects, and relevant systems. | Snowflake, dbt, Looker, Amplitude, and PostgreSQL syntax for operational queries |
| [KEY METRICS] | List the important business metrics and how they are currently understood. | trial-to-paid conversion, weekly active subscribers, 30-day retention, churn, and net revenue |
| [STAKEHOLDER AUDIENCE] | Describe who receives analysis and the decisions they make. | Growth lead, product managers, and finance manager preparing monthly planning decisions |
| [DATA GOVERNANCE RULES] | State privacy, access, retention, and review rules for using AI with data. | No names, emails, payment details, or row-level exports; use aggregated examples and approved schema text only |
How to use
- Set the tool, SQL dialect, and data-governance boundary before copying any prompt into an AI tool.
- Use the SQL and metric prompts with schema, grain, and definitions rather than asking the model to guess them.
- Run the stated reconciliation or QA query before sharing any conclusion.
- Follow up with: “Convert prompt 4 into a version for [TOOL] and add a test for [KNOWN DATA ISSUE].”
Variations
SQL debugging
Use this when a query runs but produces suspicious totals or duplicates.
Help me debug this [SQL DIALECT] query: [QUERY]. The expected grain is [EXPECTED GRAIN], the tables and join keys are [TABLE DETAILS], and the unexpected result is [SYMPTOM]. Diagnose likely row multiplication, date-boundary, null, filter, and aggregation issues. Return a corrected query with comments plus three reconciliation queries: row count by key, unmatched join count, and total comparison. Do not invent schema fields. Ask up to 3 questions only if required details are missing.
Experiment readout
Use this for a cautious analysis plan or narrative after an A/B test.
Create an experiment readout for [EXPERIMENT]. The hypothesis is [HYPOTHESIS], primary metric is [PRIMARY METRIC], guardrails are [GUARDRAILS], population and dates are [POPULATION AND DATES], and results are [RESULTS]. Produce a one-page memo with method, exposure definition, results table, uncertainty limits, segment findings, and decision recommendation. Separate observed effects from causal conclusions, flag sample-ratio mismatch or multiple-testing concerns, and do not claim significance without supplied evidence. Ask up to 3 questions only if inputs are missing.
Executive dashboard brief
Use this to turn validated dashboard changes into a decision-ready update.
Write an executive dashboard brief from [VALIDATED METRICS] for [AUDIENCE]. Compare [CURRENT PERIOD] with [COMPARISON PERIOD], using these definitions and caveats: [METRIC NOTES]. Produce five bullets: what changed, the magnitude, likely drivers supported by evidence, risks or unknowns, and the decision requested. Name base sizes where supplied and distinguish correlation from cause. Do not add explanations not in the notes. Ask up to 3 questions only if required inputs are missing.
Tips
- Always state table grain before asking for a join; most AI-generated SQL errors come from an unstated many-to-many relationship.
- Keep a metric dictionary outside the chat and paste the relevant definition when metrics drive a decision.
- Use AI to propose checks, but reconcile totals in the warehouse or BI tool that holds the source of record.
- For experiment prompts, require the exposure event, analysis window, and guardrails before allowing a recommendation.
FAQ
Can I paste production data into ChatGPT or Claude?
Only if your organization has approved that tool and data class. In many cases, schema text and aggregated results are sufficient for the prompt.
Can AI write production SQL?
It can draft SQL, but an analyst should inspect joins, filters, dialect syntax, and reconciliation results before running it in a production workflow.
How can AI help explain an anomaly?
It can structure hypotheses and suggest cuts of the data. It should not turn a pattern into a causal explanation without evidence.