# Give an AI agent boundaries before you give it tools

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

## Useful result

You'll finish with written limits and approval points for an AI agent.

An agent needs a clear job and a clear fence around it. “Be careful” is a mood, not a boundary. Name the folder, the allowed changes and the point where it must hand the work back.

## Set up

Allow 20 minutes with a tool-enabled assistant and copied sample files. Use account or tool permission controls as well as written instructions. The available controls depend on the client.

## Practice input

Task: propose new names for three files: notes-final.txt, notes-final2.txt, notes-actual-final.txt. Allowed: read the copies and create a proposed mapping. Not allowed: rename originals, delete files or share them.

## Do the task

1. Write the scope and success condition before giving the agent tools.
2. Ask for a proposed mapping only and inspect the reasoning against file contents.
3. Approve a specific next action only after checking the exact files involved.

## Prompt

Read only the three practice copies. Propose a naming table with current name, suggested name and reason. Do not rename or delete anything. Stop if you cannot tell which file is current.

## Refine

List any uncertainty about versions. Do not infer that a file is newest just because its name says actual-final.

## Expected checks

The first result must be a proposal, with no file changes. The task requires evidence about versions; the funniest filename is not evidence.

## Save

Save the proposed mapping and your approval decision. Keep originals unchanged until the mapping has been verified.

## Sources

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