A backup is a promise until someone rehearses the restore

Here is a hypothetical maintenance drill for a small AI project. A folder contains prompts, example inputs and the configuration needed to reproduce a draft. The team has a backup—but has never tried restoring it.

Give a willing teammate a clean workspace and the recovery instructions. Ask them to restore a harmless fictional example without asking the original author for undocumented steps.

Record four things: what was recovered, what was missing, which instructions required interpretation, and how you checked the result. A copied file is one observation; a reproducible output with the expected dependencies is another.

Keep secrets and real client data out of the exercise. Use synthetic inputs and the minimum configuration needed for the test. Agree beforehand who maintains the recovery notes and what upkeep they can reasonably support.

If the drill depends on one person remembering a hidden step, add that step to the instructions or reduce the promise. Preserve what is useful without keeping every old artifact forever.

Which missing dependency would be easiest to overlook: the prompt version, a source file, an environment setting, or the check that tells you recovery worked?