Skip to main content

Prompt Families, Prompts & Versioning

If an agent is an employee, a prompt is its job description — what it's actually supposed to do. Prompts are organized and versioned so you can revise what an agent does without editing workflows or re-wiring anything.

Three layers​

Prompts are modelled in three levels:

  • A prompt family groups related prompts under a theme — for example, Script writing prompts.
  • A prompt belongs to exactly one family and represents one job — Generate script, Factcheck the script, Analyze for gaps.
  • A prompt version is one specific revision of a prompt's text. Every time you edit a prompt's wording, you create a new version.

A family carries a name and a short description; a prompt carries a name and an auto-generated, unique slug.

The active version​

A prompt can have many versions over time, but exactly one is active at a time — recorded on the prompt itself as its active version. When a workflow runs, it uses whichever version is active at that moment. Changing which version is active is how you roll a change out — or roll it back: promote a new version to try it, and if it's worse, re-activate the previous one. Activating a version affects future runs only; it never alters a run already in progress.

The safety this buys​

Because the wording lives in versioned prompts and the active pointer is a single switch, you can iterate on what your agents do — confidently and reversibly — without ever touching a workflow or an agent. And like the rest of the configuration chain, disabling a prompt family cascades down through its prompts to the configurations and workflows that use them, so a disabled prompt can't silently leave a workflow half-broken.