Delegate a Task to an AI Agent Safely

Prompt

Task briefing. Follow it exactly and stay inside the boundaries.

Goal: {{goal}}
Done means: {{definition_of_done}}

You may work with: {{allowed_scope}}
You must not touch anything else, and specifically not: {{forbidden}}

Limits: stop after {{max_steps}} steps, or sooner if you are repeating yourself.

Stop and ask me if: an action cannot be undone, something is outside the scope above, the instructions turn out to be ambiguous, or you have failed the same step twice.

When you finish, report: what you did step by step, what you assumed, what you could not verify, and anything you changed that I might not expect.

Fill in these placeholders

• {{goal}}
• {{definition_of_done}}
• {{allowed_scope}}
• {{forbidden}}
• {{max_steps}}

How to use it

Fill in the scope and the forbidden list properly, in specifics rather than categories. Naming the exact folder, label, or file is what keeps the work contained; naming a general area is not.

Example use case

Asking an assistant to tidy one project folder while making clear it must not touch anything outside it.

Why this prompt works

Most agent trouble is scope creep rather than incompetence: given an open goal, it does more than you meant, and every individual decision looks defensible. Stating what done means removes the incentive to keep going, and the stop conditions convert the moments where judgment is genuinely needed into a question rather than a guess.

Common mistakes

• Describing the scope in general terms. 'The project files' invites interpretation; a named folder does not.
• Leaving out a definition of done, which is how tidying a document becomes rewriting it.
• Treating the briefing as a substitute for permissions. What the software allows matters more than what the briefing asks.

Variations worth trying

• Add 'do not create any new files' when you only want existing ones changed.
• Ask for a dry run first: describe every action without performing any of them.
• Lower the step limit deliberately for a first attempt, then extend it once you have seen how it works.