You’re about to manage AI agents, not just people. Here are the 5 habits that make you great at it, with an exercise for each one you can do this week.
As agents take on more of the work, your job turns into deciding what to hand off, giving direction and judging whether the result is any good. That takes practice.
Each habit below has one exercise to do this week in Claude or ChatGPT, on a real task from your own work. Do one a day.
For the bigger picture on the role, read my AI agent manager guide.
Without clear limits, you either approve every tiny step and save no time, or the agent makes a call you’d never make, like offering a discount. Good managers say up front what it decides alone and where it stops.
01
Write your decision-rights note
Pick one job you hand AI every week, like customer emails or social posts. Paste this and answer its questions.
02
Save it where it gets read
In Claude, open Projects, click + New Project, then Set project instructions. Paste the note and click Save instructions. Every chat in that Project starts with it. In ChatGPT, open your project’s three-dot menu and pick Project settings.
Claude’s Auto setting lets it work without asking about each step, so your note is what tells it where to stop.
What Good Looks Like
A new hire could follow every line without asking you. “Refunds up to $50 are yours” passes. “Use good judgment” fails.
Most bad results start as a gap you left that the agent filled with a guess. Asking for the plan and the missing pieces first catches the wrong turn on paper, not after an hour of wasted work.
01
Ask before it starts
Pick the biggest task on your list this week and paste this first.
Using Claude Code? Plan mode does this for you. My delegation power moves guide shows how.
What Good Looks Like
You fixed at least one wrong assumption before any work started. If it asked nothing, push back with “What would you have to assume?”
One agent doing one task while you watch is still you doing the work. Time comes back when two jobs run at once and you only step in to review. The skill is splitting work into pieces that don’t depend on each other.
01
Pick two tasks that don’t touch
Good pairs are competitor research plus a newsletter draft, or cleaning a customer list plus writing FAQ answers. If task two needs the result of task one, it’s not a pair.
02
Start both at the same time
If your message box no longer shows separate Chat and Cowork options, you have Claude’s combined experience. Send the first brief below, click New chat, send the second. Both run at once, longer tasks keep going in the cloud with your laptop closed, and both show up in Recents.
Still see separate options? Use two browser tabs, one brief each. Same in ChatGPT.
Longer list? Claude can split it across sub-agents itself.
What Good Looks Like
Both came back without you touching either chat. If one stopped to ask you something, add that answer to your brief for next time.
“That’s wrong, fix it” fixes one answer. “That’s wrong because my customers are retirees and slang turns them off” fixes every answer after it. Saving it stops you giving the same feedback every week.
01
Correct with the reason
Next time you catch a mistake on real work, paste this.
02
Save it so it sticks
One rule? Paste it into your project instructions next to your decision note.
A whole weekly job? Say yes to the skill offer and let Claude build it. Click Save skill on the skill it hands you (some accounts say Copy to your skills). Only got a download? Go to Customize (being renamed Plugins, so you may see either), then Skills, click +, then Create skill, then Upload a skill.
Skills work on every plan once Code execution and file creation is on in Settings, then Capabilities (newer accounts call it Create and edit files). More ways are in my stop AI repeating mistakes guide, and my Lock It In skill turns a chat that went well into a tested skill.
What Good Looks Like
In a fresh chat in that Project, the same kind of task comes back without the mistake.
When an agent does 200 of something, one small misunderstanding becomes 200 mistakes. Reviewing 2 first takes five minutes, and your corrections become the standard for the other 198.
01
Approve 2, then let it run
Use a real batch of 20 or more, like product descriptions or follow-up emails. Attach the list and paste this.
What Good Looks Like
You’d publish both samples as is, and 5 random spot-checks from the rest are clean too.
On Friday, reopen your decision note and move one item from “Check with me first” to “Decide on your own.” That’s how trust grows.