Natural-language-to-data-visualization

The task

You want a model to turn data and intent into a chart, dashboard element, or visualization spec. The hardest part is usually not drawing the chart. It is choosing the right comparison and avoiding misleading encoding.

Why the old approach is outdated

“Make a chart from this data” lets the model choose the story without knowing the decision. It may pick a chart type that looks normal but hides the real comparison.

The current approach

Recommend a chart for this data.

Decision: decide whether support volume changed after the pricing-page release.
Data: daily ticket count, category, plan tier, and date.
Constraints: avoid dual-axis charts, show uncertainty if the sample is small, and do
not claim causation from this data alone.

Return the chart type, required fields, aggregation, and one sentence the chart can
support without overstating the data.

If the visualization spec will be consumed by code or a charting service, pair the prompt with structured output prompting instead of relying on free-form prose.

When to use a different model

Use Gemini when the chart is created through a multimodal or tool-calling workflow. Use Claude when the work is mostly critique of an existing chart or memo. Use ChatGPT when you also need code, a spreadsheet formula, or a structured visualization spec.

What to avoid

• Asking for a beautiful chart before asking for the correct comparison.
• Letting the model infer causation from adjacent dates.
• Skipping units, date ranges, and missing-value rules.
• Accepting generated chart code without checking the data transformation.