What should we keep from an AI experiment besides the successful output?
A result one enthusiastic colleague gets can be hard for anyone else to repeat. Keep the task, approved input, relevant configuration, evaluation criteria, version information and test date so someone else can run it again, and note every correction and manual step along the way.
Use whatever version information the product exposes, and keep secrets and confidential material out of a casual shared notebook. If the final output depended on several expert interventions, those belong in the account of how it was produced.
Then ask another permitted participant to repeat the task. Compare their result against the same standard and note where the experience differs. Identical wording isn't required if the task allows several good answers.
The record helps you decide whether the approach is teachable and maintainable. No need to preserve every click. What you're guarding against is a successful example turning into a mysterious recipe that only its original creator can operate.