Multi-step agent task plan with checkpoints and stop rules

Before taking any action, write a plan for the task below and wait for my approval.

Task: [WHAT YOU WANT DONE, e.g. "update the dependency versions in this repo and fix any failing tests"]
Definition of done: [HOW WE KNOW IT IS FINISHED, e.g. "all tests pass and a summary of changes is written"]
Tools and access you have: [e.g. "file system in this folder, terminal, web browsing", OR "list what you can use"]
Constraints: [e.g. "do not push to remote", "budget: 30 minutes", "do not contact anyone"]
Context: [FILES, LINKS OR BACKGROUND]

The plan must include:
1. Restated goal and done criteria, plus any assumptions. List questions if the task is ambiguous.
2. Numbered steps. For each: action, tool, expected result, and how you will verify it worked.
3. Risk labels: mark any step that deletes, overwrites, spends money, sends messages, or changes anything outside [SCOPE] as "NEEDS APPROVAL".
4. Checkpoints: after which steps you will pause and report progress.
5. Stop rules: stop and ask me if a step fails twice, if results contradict the plan, if you need access you do not have, or if the task is larger than expected.
6. A rollback note: how to undo each change.

After approval, follow the plan, report at each checkpoint, and if you need to deviate, explain why before doing it.

Example (excerpt):
Goal: update dependencies to the latest minor versions; done when the test suite passes.
Assumption: major version upgrades are out of scope.
1. Read package.json and the lockfile (file read). Expected: list of outdated packages. Verify: compare with "npm outdated".
2. Create a new branch "deps-update" (terminal). Rollback: delete the branch.
3. Update minor versions and run tests (terminal). Checkpoint: report failures before fixing.
4. NEEDS APPROVAL - modify test files if a failure is caused by a changed API.
Stop rule: if more than 5 tests fail after step 3, stop and report.

When to use it

Use this prompt with any AI agent that can take actions on its own: coding agents, browser agents, or automation tools connected to email, files or business systems. It is most valuable for tasks with several steps, real side effects, or unclear scope, where a wrong first move is expensive.

For a one-step task, such as renaming a variable or summarizing a single file, the plan is overhead and you can skip it. The value grows with the number of steps and with how hard the actions are to undo. Many agent tools also offer a built-in planning or approval mode; this prompt works alongside those features and gives you a consistent format for the plan you review.

How to customize it

• [TASK] should describe the outcome, not the method; let the plan propose the method.
• [DEFINITION OF DONE] prevents the agent from stopping too early or wandering into extra work.
• [TOOLS AND ACCESS] should match what the agent really has. If you are unsure, ask it to list its tools first.
• [CONSTRAINTS] and [SCOPE] define the boundary that triggers “NEEDS APPROVAL”.
• [CONTEXT] reduces guesswork; point to specific files or documents.

Why it works

Agents tend to fail by acting on a wrong assumption and then compounding it over many steps. A written plan surfaces assumptions before anything changes, and reviewing it takes a minute compared with untangling a long run.

Verification for each step, explicit risk labels and stop rules turn a vague request into a controlled process. Checkpoints limit how far the agent can drift, and the rollback note makes recovery possible. This does not replace technical safeguards such as limited permissions, but it makes the agent’s intentions visible.

Tips and variations

• Combine the plan with least-privilege access; see our guide to API keys for coding agents.
• For coding agents, pair this with a standing system prompt for a careful coding agent.
• Ask for a final report that compares what was planned with what was done.
• Read how to evaluate AI agents before automating a recurring workflow.