Data analysis prompt builder
LLMs can reason about tabular data surprisingly well, but a vague “analyze this CSV” produces hand-wavy claims and invented numbers. The data analysis prompt builder assembles a prompt that gives the model your column schema, a few sample rows, and a clear goal, wrapped in rules that demand statistical rigor and honest caveats — so the output is grounded rather than guessed.
How it works
You describe your columns and their types, paste a handful of representative rows, and state your analysis goal. The tool composes a prompt that asks the model to state its assumptions, quantify every finding with concrete numbers, keep correlation distinct from causation, and explicitly flag missing values, outliers, and small-sample limits. Where a visualization would help, it asks for a precise chart description — type, axes, grouping, and insight — instead of a drawing. You choose the output format (narrative, bullets, metrics table, or JSON), and the whole thing runs locally with no API key.
What makes a data analysis prompt rigorous
The default LLM response to “analyze this CSV” is confident, fluent, and often wrong in ways that are hard to spot — invented correlations, wrong aggregations, missing caveats about small samples. The generated prompt addresses these failure modes directly:
Assumption statement. The model lists what it assumes about the data before analyzing (e.g., “assuming Date is in ISO format, Region has no duplicates due to casing”). Stated assumptions are easy to correct; silent ones cause invisible errors.
Quantified findings. The prompt forbids phrases like “revenue was high” and requires concrete numbers. If a numeric claim cannot be supported from the sample, the model must say so.
Correlation vs. causation. The prompt includes an explicit rule: “describe correlations only; do not assert causal relationships without noting confounders.” This prevents the model from writing “X causes Y” when all it observed was co-occurrence.
Small-sample caveats. With only a few sample rows, many statistical claims cannot be made. The prompt asks the model to flag where the sample size limits confidence — something models almost never do unprompted.
Choosing the right output format
| Format | Best when |
|---|---|
| Narrative report | Presenting to non-technical stakeholders |
| Grouped bullets | Quick review by a data team |
| Metrics table | Feeding results into a dashboard or doc |
| JSON | Piping analysis into another tool or pipeline |
The chart description format — regardless of which you choose — requests chart type, axes, groupings, and the insight each chart would show, so you can build the visualization in any tool without ambiguity.
Tips and notes
- Give honest sample rows. Include an edge case or two — a null, an outlier — so the model sees the messy reality, not just clean examples.
- State a specific goal. “Which region drives revenue and where did it dip” beats “tell me about the data” by a mile.
- Trust numbers, verify them. The prompt demands quantified claims; still spot-check the arithmetic against your real dataset.
- Use JSON for pipelines. The suggested-charts array drops straight into a visualization step or report generator.