Give an AI agent a task without letting it run away

You're going to work on this on your own for a while. Before you start a single step, write out your plan and your stopping rules, and wait for me to approve them.

The job: [WHAT YOU WANT DONE, AS AN OUTCOME — not a list of steps]

What "done" looks like: [HOW YOU'LL KNOW IT WORKED]

Your plan must include:
1. The steps you intend to take, in order.
2. What you'll do FIRST that would prove the approach is wrong if it fails — cheapest disproof first, not easiest step first.
3. Your budget: how many steps or how long before you stop and report back regardless of progress.
4. STOP AND ASK ME conditions. Write these as specific triggers, not good intentions. At minimum: anything that sends something to another person, anything that spends money, anything that deletes or overwrites, anything that can't be undone, and anything where you've tried the same thing twice and it failed twice.
5. What you will NOT touch, even if it looks like it would help.

While you work:
- Report at each checkpoint in three lines: what you did, what you learned, what's next.
- If you're stuck, say "stuck" and describe what you tried. Do not quietly switch to a different approach and keep going.
- If you're guessing, label it a guess in the moment, not in a summary at the end.

Don't start until I've approved the plan. If my job description is ambiguous in a way that changes your plan, ask me now rather than picking one reading.

One more thing: use anything you already know about me from our past chats and anything you're connected to. If you can find it yourself, don't make me look it up.