Pradra StudioDocs
Sign in

Prompts

For a generative capability, the prompt is the specification: it tells the model what to do with each request. Pradra Studio treats prompts the way it treats models — versioned, evaluated, and gated — because changing a prompt changes production behavior just as surely as swapping a model.

How prompts work

A prompt is a template with placeholders for the capability's declared inputs:

Summarize the following support ticket in three sentences,
preserving any order numbers exactly.

Ticket: {{ticket_text}}
Customer tier: {{tier}}
  • Placeholders are validated at save time: every {{field}} must exist in the capability's input contract, so a typo is caught when you write it, not when a customer hits it.
  • At serving time, user input only fills the holes. The instructions around them are yours alone — a request cannot rewrite the prompt from the inside. This separation is the platform's first defense against prompt injection.

Versions

Every edit creates a new prompt version; the version that is serving stays untouched until you explicitly activate another. The Studio shows a diff between versions, so a review is "what changed?" — not "read both and guess".

The activation gate

Activating a prompt version is a promotion, and it passes through the same kind of gate as a model:

  • If the capability carries acceptance criteria over its generative metrics, a version with no successful evaluation cannot activate — an unevaluated prompt can't reach production.
  • A version whose golden-set score fails the criteria is blocked, with the failing rule named.
  • Rolling back is activating the previous version — instant, like every rollback in the platform.

Working with prompts in the Studio

  1. Open the generative capability in AI Studio and go to its Prompts tab.
  2. Write the template; save creates version 1. Edit again — version 2, with a visible diff.
  3. Try a version against real inputs in the Playground tab before evaluating.
  4. Run an evaluation, then Activate the version that passed.

Writing prompts that hold up

  • Be explicit about format. "Three sentences", "JSON with these keys", "preserve order numbers exactly" — vague instructions produce confident, vague output.
  • Put rules before data. State the instructions, then introduce the user-filled fields.
  • Change one thing per version. Evaluation tells you whether a version is better; small diffs tell you why.

Next