Make the AI Plan Before It Acts

Prompt

Before doing anything, write your plan for this task and then stop.

Goal: {{goal}}
What you can access: {{available_tools}}

Give me:
1. A numbered list of the steps you intend to take, with the specific action or tool for each one.
2. For each step, what could go wrong and how you would notice.
3. Which steps cannot be undone once done.
4. Anything you are unsure about or assuming.

Do not begin any step, and do not take any action that sends, deletes, publishes, or pays, until I reply with approval.

Fill in these placeholders

• {{goal}}
• {{available_tools}}

How to use it

Use this as the first message of any agent run that touches real files, accounts, or messages. Read the plan properly. The five minutes it takes is the cheapest point at which a misunderstanding can be corrected.

Example use case

Asking an assistant to reorganize a shared folder, and discovering from its plan that it intended to rename files you rely on.

Why this prompt works

The instruction to stop is doing the work. Left alone, a capable agent starts acting immediately, and the first thing you see is a result rather than an intention. Asking for irreversible steps to be named separately also forces the model to sort its own plan by consequence, which surfaces the risky part in a place you cannot skim past.

Common mistakes

• Leaving the available access vague, so the plan is written against tools it may not have and has to be redone.
• Skimming the plan and approving it. An unread plan provides exactly the same protection as no plan.
• Approving the whole sequence at once when only the first few steps were actually clear.

Variations worth trying

• Add 'give me the shortest plan that achieves this' to see whether half the steps were optional.
• Ask what it would need access to that it does not currently have, which often reveals the task is not what you thought.
• Ask for the plan twice in fresh conversations and compare. Divergence tells you the goal is underspecified.