Skip to content
Agentivity
Star on GitHub Start free

Try, test and improve a prompt

Prompt Studio has four ways to work on a prompt: try it, test it, analyse it and optimise it. All four send the prompt to an AI model.

All of this needs a model connection chosen in the tab Settings of the panel on the right.

Each execution uses the AI model, and your provider bills you for it.

  1. In the tab Playground, give a value to each input.
  2. Select Run Prompt.
  3. Read the answer in the Debug console, at the bottom.

Playground history keeps the list of your tries. Select one to see its output again.

A test case is one example that you keep: values for the inputs, and the answer you expect. It lets you check the prompt again after each change.

  1. Select the tab Test Cases.
  2. Select Add Test Case.
  3. Give it a Name.
  4. Fill in the inputs.
  5. Fill in the expected result. It is required: it is the reference used to check the answer.
  6. Select Save.

The tab Test Cases of Prompt Studio: the buttons Run All Test Cases, Refresh and Add Test Case, an empty list of test cases, and the history at the bottom

  • Run All Test Cases runs every saved case against the current prompt.
  • Select one case to see its last execution: the expected answer, the actual answer, and a score.

A case that passes is marked Success; one that does not is marked Failed, with an analysis of the failure.

Use test cases for what must not break: after each change of the prompt, run them all again.

  1. In the tab Playground, select Analyze.
  2. Read the Analysis Result.
PartWhat it tells you
StrengthsWhat the prompt does well.
IssuesWhat is wrong, each with a level: HIGH, MED or LOW.
RecommendationsWhat to change.
Improved PromptA rewritten version of your prompt.

To take the improved version, select Apply. It replaces the text of your draft. It is saved only when you select Save.

Optimisation tries several versions of the prompt in a row and keeps the one with the best score.

  1. Open Optimization.
  2. Set Max Iterations: how many versions it may try.
  3. Select Start Optimization.
  4. When it is finished, read Score per Iteration and the Best Candidate.
  5. Select Apply Best Candidate to take it, then Save.

Optimisation makes many calls to the AI model. Start with a small number of iterations.

The scores are given by the AI model too, following an evaluation prompt. The last tab of the panel on the right lets you write your own. Leave it empty to use the default evaluation.

  • A score is an opinion of an AI model. Read the answers yourself before you trust a number.
  • Nothing is changed without you: an improved prompt is applied only when you select Apply, and saved only when you select Save.

Back to What Prompt Studio is for.