Prompts and versioning
A test case can carry its system prompt as free text, or link to an entry in the Prompts library. Linking means the text lives in one place and every case that references it moves together.
Each prompt keeps its full history. There is no “new version” button: saving changed content appends a version, and the old one stays. Renaming a prompt or editing its description does not, since neither changes what the model sees.
Pinned or latest
Section titled “Pinned or latest”A linked case stores the prompt id and a version, and that version is either a fixed number
or latest.
Latest follows the prompt. Edit the prompt and every case set to latest picks up the new text on the next run, which is what you want while you are still iterating on wording.
Pinned freezes the case on one version. The prompt can move on and the case will not. Pin when a case exists to catch a regression and you need its instructions held still, or when you want two cases to run different versions of the same prompt side by side so the suite compares them directly.
A pinned case shows its prompt read-only. Editing means switching it back to latest, which also swaps in the latest text.
Editing from a test case
Section titled “Editing from a test case”The system prompt field in the test case editor is a live editing surface for the prompt itself when the case is set to latest. Changing the text there and saving the case appends a version to the shared prompt, so other cases on latest are affected too.
Free text that is not linked to anything can be promoted with the bookmark button next to the field. Cannonade creates the prompt and switches the case to it.
Restoring an old version
Section titled “Restoring an old version”Selecting an earlier version in Prompts shows it read-only. Set as Latest copies its content forward as a new version rather than deleting anything after it, so the history stays append-only and cases pinned to the versions in between keep working.
What runs actually record
Section titled “What runs actually record”A run resolves every prompt reference the moment it starts and stores the resulting text in the run file. The run is a record of what was sent, not a pointer to what the prompt says today, so later edits never rewrite past results.
That is what turns prompt engineering into something you can measure rather than guess at. Change the prompt, run the same suite again, and the two runs each hold their own version of the instructions, so an improvement you think you made is either in the scores or it is not.
Sharing
Section titled “Sharing”Prompts are individual JSON files, separate from suites. A suite that links to prompts is not self-contained: copying it to another machine without the matching prompt files leaves those cases pointing at ids that do not resolve. Copy both, or keep the prompt text inline in cases you intend to hand around on their own.