# What your team needs before an AI agent touches real work

Original 5HO practice material. Fictional example. Reviewed 5 October 2026.

## Useful result

You'll finish with a boundary checklist with a named shutdown owner.

Before an agent touches live work, give it a smaller place to practise. A successful demonstration is useful. A clear recovery plan is useful on the day the demonstration stops being successful.

## Set up

Allow 30 minutes to prepare a sandbox: a separate folder or test workspace. You need an agent with visible tool permissions and an owner who can stop it.

## Practice input

Task: organise 8 copied training files into Notes, Purchases and Archive. Allowed: read and move copies inside the test folder. Forbidden: delete originals, send messages or change sharing permissions.

## Do the task

1. Record the allowed folder, actions and limits outside the prompt.
2. Inspect the actual tool permissions and restrict them where the product permits.
3. Run a small sample and check both the output and how to restore the originals.

## Prompt

Draft a readiness checklist for this file-sorting agent. Include test copies, allowed actions, forbidden actions, review point, move log and recovery steps. Identify permissions I must configure myself.

## Refine

Add stopping conditions for access outside the test folder, unexplained deletions or an unavailable restore copy.

## Expected checks

A prompt saying “do not delete” is not a permission setting. You should know the exact test location and be able to account for all eight files before moving to real work.

## Save

Save the checklist with the test result and restore location. Expand scope only after the owner reviews it.

## Sources

- Tool permission controls in Claude Code: https://code.claude.com/docs/en/permissions