← Back to all free guides
A free guide by Mariah Brunner ↗ INSTAGRAM ↗ TIKTOK

Give Your AI Agents
These 4 Rules

If you spend more time supervising your agents than they save you, these four rules fix it. Each one comes with a line you can paste.

Anthropic’s own Claude Code guide puts the problem perfectly: if your agent has no way to check its work, "you’re the test suite." These rules hand that job back to the agent. They work in Claude, ChatGPT or any agent you use.

The 4 Rules

01

Describe the result, not the steps

Tell it what finished looks like and what limits to follow, then let it work out how. Asking it to say how it will check the work first catches a misunderstanding before it spends twenty minutes on the wrong thing.

Brief Your Agent
Here's what I want done: [the finished result in one or two sentences]. It's done when [what finished looks like, like "a 1-page doc in my Drive with 5 ranked ideas"]. Stick to these limits: [budget, tone, files you can touch, what not to change]. You figure out the steps. Before starting, list how you'll check that this worked.

02

Make it prove the work is done

Agents stop when the work looks done. Ask for proof, not a promise. Add this to the end of any big task.

Add This At The End
When you finish, show what you tested, the results and anything you couldn't complete or verify. Don't just tell me it works. Show me how you know.

03

Give it work and walk away

Think as big as possible and hand over the whole job. Just tell it when to come back to you.

Tell It When To Come Back
Work on this without checking in with me. If you've tried three times and you're still stuck, stop and explain what's blocking you and what you need from me. Ask before you delete, send or pay for anything.

In Claude, tasks now run in the cloud on Pro and Max, so you can close your laptop and Claude keeps going. Pick Auto in the permission menu so it doesn’t stop to ask about every step. It pings your phone when it’s done or needs you.

04

Schedule the work you keep repeating

Not sure what to automate? Ask your agent. This finds the work you do on repeat and suggests what to put on a schedule.

What Can You Automate For Me?
Look through my recent chats, files and connected apps and find the work I keep repeating. Suggest 3 to 5 tasks you could run for me on a schedule. For each one, tell me what it does, how often it should run and what you'd need access to. Start with the one that saves me the most time.
  1. In Claude: type the task and how often, like "Every weekday at 8am, give me the 5 emails I need to answer." Answer its questions, check the details and click Schedule. Not seeing that yet? In Cowork, open Scheduled in the left sidebar, then New task. It runs hourly, daily, on weekdays or weekly, on any paid plan. (Claude’s steps)
  2. In ChatGPT: type the task and when in a chat, then confirm the schedule it suggests. Results come by notification or email.
  3. Watch the first few runs before you trust it on its own.
Paste These Rules In Once

Put this in your Claude Project instructions (or your CLAUDE.md file in Claude Code) so every task starts with all four rules.

Always-On Agent Rules
## How to work for me - Start from the result I described. Work out the steps yourself. - Before you start, list how you'll check the work is done. - When you finish, show what you tested, the results and anything you couldn't verify. - If you've tried three times and you're still stuck, stop and tell me what's blocking you. - Ask before you delete, send, publish or spend anything. - If you notice I repeat a task, suggest scheduling it.

Want your agent to check and retry its own work in a loop? Here’s the difference between prompts, loops and graphs.