Data Analysis ChatGPT Prompts
Data interpretation, analytics, and insights generation prompts
Create a Data Cleaning Procedure
Write a repeatable data cleaning procedure for {dataset_type} from {source}. Cover in order: profiling before changing anything (row counts, distributions, null rates, cardinality), schema and type validation, duplicate detection and the matching rule for near-duplicates, missing value handling with the decision rule for drop versus impute versus flag — and the bias each introduces, outlier detection separating genuine extremes from errors, standardising formats (dates, currencies, categoricals, text casing and whitespace), and referential integrity. For every step, log what changed and how many rows were affected, so the cleaning is auditable and reversible. Add the validation suite that runs afterwards and the thresholds that should halt the pipeline rather than passing degraded data downstream.
Write a Statistical Analysis Interpretation
Interpret the results of {analysis_type} on {dataset} for a non-technical audience making {decision}. Translate each statistic into a plain sentence about the real world, then state the practical significance separately from the statistical significance — a result can be highly significant and commercially irrelevant. Give the effect size with its confidence interval in units the audience uses. State the assumptions the method requires, whether they hold here, and what it means for the conclusion if they do not. List the alternative explanations you cannot rule out with this data, and the specific additional data that would settle it. End with what you would and would not decide on this basis.
Design a Customer Segmentation Analysis
Design a customer segmentation for a {business_type} with {customer_count} customers, to be used for {intended_use}. Choose the approach that fits the use — RFM for retention and lifecycle, behavioural for product decisions, value-based for prioritisation, needs-based for messaging — and justify it rather than defaulting to k-means. Specify the features, the treatment of outliers and missing data, and how the number of segments is chosen and validated. For each resulting segment give: size, share of revenue, defining characteristics, and the specific action it enables. Apply the usability test — a segmentation nobody can act on is an academic exercise — and add the refresh cadence and the drift signals that mean it needs rebuilding.
Create a Predictive Model Framework
Design a predictive modelling framework for {prediction_target} at a {business_type}. Start with the framing: what decision the prediction changes, the cost of a false positive versus a false negative, and the baseline (often a simple rule) that any model must beat to justify its complexity. Then cover: target definition and the prediction window, feature engineering with explicit leakage checks, a train/validation/test split that respects time ordering, model selection starting simple, and evaluation using metrics chosen for the business cost rather than accuracy — which misleads on imbalanced targets. Add calibration, interpretability requirements, the deployment and monitoring plan including drift detection and a retraining trigger, and the fairness review where the prediction affects people.
Write a Data Governance Policy
Draft a data governance policy for a {company_size} {industry} organisation. Cover: data ownership and stewardship with named roles and real accountability, a classification scheme (public, internal, confidential, restricted) with the handling requirement for each, access control and the request and review process, retention and deletion schedules by data type, quality standards and who enforces them, a business glossary resolving the definitional disputes that make two reports disagree, lineage documentation requirements, and privacy obligations under the applicable regimes. Include the governance forum, its decision rights, and an exception process — because a policy with no legitimate exception path gets routed around entirely. Keep it operational rather than aspirational.
Develop a Data Quality Assessment
Design a data quality assessment for {data_domain} at a {company_type}. Assess across six dimensions with a concrete measure for each: completeness, accuracy against a trusted reference, consistency across systems, timeliness relative to the decisions it supports, validity against business rules, and uniqueness. For each, define the test, the acceptable threshold justified by downstream use rather than an arbitrary percentage, and the measurement frequency. Produce a scorecard by source system, a prioritised issue list ranked by downstream business impact, and root-cause categories distinguishing collection problems from pipeline problems from definitional disagreements. Add the ongoing monitoring and the alerting rules that catch degradation before a report is published on bad data.
Create an A/B Test Analysis Template
Build an A/B test analysis template for {test_type} experiments on {metric}. Include the pre-registration section completed BEFORE launch: hypothesis, primary metric, guardrail metrics, minimum detectable effect, required sample size with the power calculation, and planned duration covering full business cycles. Then the analysis: results with confidence intervals rather than point estimates alone, the p-value with a plain-language interpretation, guardrail checks, and pre-declared segment analysis. Add the validity checks most teams skip — sample ratio mismatch, novelty effects, and whether the test actually ran as configured. Finish with a decision rule agreed in advance, and explicit guidance on why stopping early at significance inflates false positives.
Create a Data Analysis Report
Write a data analysis report on {dataset} answering {business_question} for {audience}. Lead with the answer and the decision it supports — not the methodology. Then: the three findings that matter with the numbers and their uncertainty, the method summarised in a paragraph with detail in an appendix, the data quality issues and how they constrain the conclusion, what the data cannot tell you, and the recommended actions with the expected effect of each. Every chart must have a title that states its finding rather than naming its axes. Distinguish correlation from causation explicitly wherever a reader might over-read. Keep the main report to two pages.
Write Data Visualization Recommendations
Recommend visualisations for presenting {data_type} to {audience} to support {message}. For each recommendation state the chart type, why it suits this comparison (change over time, part-to-whole, distribution, correlation, ranking), and the alternative you rejected and why. Cover the craft: axis handling including when a non-zero baseline is defensible and when it misleads, colour used to encode meaning rather than decoration, accessible palettes tested for colour-vision deficiency, direct labelling in place of legends where possible, and a title that states the finding. Name the specific chart types to avoid for this data and the distortions they introduce. Add guidance for small multiples versus one dense chart, and how the choice changes for a live dashboard versus a static report.
Develop a KPI Dashboard Specification
Specify a KPI dashboard for {audience} at a {business_type} supporting {decisions}. Start from the decisions the viewer makes, then work back to the minimum metrics that inform them — most dashboards fail by inverting this. For each metric define: the precise calculation, the source system, the grain, the refresh frequency, the comparison basis (versus target, prior period, or same period last year), and the threshold that should prompt action. Separate the three or four headline metrics from supporting detail, and include leading indicators rather than only lagging outcomes. Specify layout priority, the chart type per metric with the reasoning, and the annotation mechanism for explaining anomalies. Add what deliberately does NOT belong on this dashboard.