How to prompt an AI for the chart you actually want

Editing a chart by chat is faster than hunting through menus, but only if the instruction is precise. Vague prompts produce charts that are technically responsive and practically useless.

Say the message, not just the shape

Make a chart of this data gives the model no basis for a choice. Show that most of our revenue is eaten by cost of goods gives it one, and the chart type follows from the sentence.

Name the elements you want changed

Make it nicer is unactionable. Colour the cost nodes red and move the Net profit label to the right names the objects and the operation, and it lands the first time.

One instruction per message

Batched requests — recolour, rename, add a node, change the title — usually get partially applied, and then you cannot tell which part failed. Small steps are faster overall.

Give the constraints up front

Amounts in thousands, at most eight nodes, colourblind-safe palette. Constraints stated early stop the loop where you keep undoing the same unwanted default.

  • Weak: make this look better.
  • Strong: group everything under 2% into Other and sort descending.
  • Weak: add the numbers.
  • Strong: show values in € thousands with no decimals on every node.
Try it in the editor

The mental model that helps: you are not asking for a chart, you are dictating edits to a colleague who can see the same canvas you can.

Frequently asked

What makes a good chart prompt?

A stated message, named elements, one operation per message and explicit constraints on units, node count and palette.

Can AI pick the right chart type for me?

It can suggest one if you describe the conclusion you want the reader to reach. Given only a data table, any suggestion is a guess.