Last week Artur asked me a question a lot of you have been asking too: how do you set up your AI as a real chief of staff? Not a clever assistant, but an actual one?
A few months ago I told you to give your AI a name, a memory, and a job. That was the setup. Now we run it.
What a real job needs
There are three things that make it a real job.
1. A job description
Most AI setups do have a job description. But there are two ways to get it wrong.
The first is being too vague: "You are my chief of staff. Help me stay focused." The AI has to guess what you actually want, and you get a different version of generic every day.
The second is the opposite: stuffing the role definition with 4,000 words of every scenario, every tone, every edge case. That isn't a job description. That's the entire onboarding manual crammed into the first paragraph.
A real job description holds three things: the role, the scope, and the boundary. Something like:
Help me decide what to work on each week and what to drop. Use the tools I've connected. Push back when my plan doesn't fit the days I have. Don't write the work for me. I want decisions, not drafts.
Keep it small. Every extra word is one more thing the AI has to weigh.
And don't write it from scratch. Ask your AI to draft its own. Give it the role and the outcomes you want, then sharpen what comes back. The first draft will tell you what your AI thinks the role is, which is useful information either way.
Everything else (the scenarios, the examples, the rules you want it to follow) lives in the files behind the chat. Same as Issue 1's setup. CLAUDE.md is the table of contents, and the details live in the folders. Your AI reads what it needs when it needs it.
The same goes for what you connect. Only give your AI access to what it actually needs. More files don't help. They just bury the ones that matter.
For a chief-of-staff role, the basics are: your calendar (where your meetings live), your docs (Google Drive, Notion, wherever you write), your tasks (whatever to-do list you use), and your team chat (Slack, email, whatever). Skip everything else.
2. A schedule
A real employee doesn't only show up when you call on them. They're around. They share context as the week goes. They ask a question when something seems off.
Your AI employee can do the same, but only if you make showing up a routine. The ritual is one daily check-in, every weekday morning. On Monday I expand it into a priorities review for the week ahead. And on Friday I expand it into a post-mortem for the week behind.
For each of these, be specific about what "brief me" actually means. Spell out what to pull from each source, what time window to look at (today, this week), what to summarize versus what to flag, and how long the output should be. Same rule as the job description: specific enough to give signals, flexible enough to leave room for the AI to think.
Daily: the check-in. Three lines work for me. Where I am with things. What's getting in my way. What I'm putting off. And you'd be surprised at what comes back when you tell your AI what you're putting off. Mine has caught me three weeks in a row on the same project I keep avoiding. It doesn't moralize. It just notices.
The exact questions don't matter much. Pick whichever three keep you honest with yourself. The shape matters more than the words.
Monday: the priorities review. Same check-in, plus one extra question: "What should I focus on this week, and what should I drop?" Most people ask what to focus on and get a list. A real partner is paid to also tell you what to ignore.
Friday: the post-mortem. Same check-in, plus a different ask. Have your AI write you a recap of the week. What got done, what didn't, what it noticed about how you worked. Save it. Paste it into next Monday's review. Without a post-mortem, every week starts cold. With one, your AI starts Monday with last week's context, not a blank slate.
3. Permissions and limits
A real employee doesn't just say yes. They tell you when your priorities don't fit. They ask the second question. They flag what you're missing. Your AI won't do any of that without permission. It'll keep saying yes.
So write the rules into the setup. Two kinds matter: when to speak up, and when to act.
Speaking up. Here are three examples (more in the Issue 1 prompt):
Stay objective. You are not a yes-bot. You are a thinking partner.
Ask before you act on anything big.
When you call me out on a pattern, use my own words back at me.
Acting. Rate the actions your AI can take by how much they matter:
Low risk: brainstorming options, suggesting tradeoffs, outlining a draft. → Just act.
Medium risk: rewriting a doc, prepping talking points, reorganizing my notes. → Do it, then tell me what you did.
High risk: anything irreversible. Sending the email. Spending money. Reaching out to someone on my behalf. → Show me first.
When it gets things wrong (the part that makes everything else compound)
If you only get one thing right from this article, get this one.
Even with the right job description, the right cadence, and the right rules, your AI is going to mess up. It's going to mess up a lot. But that isn't the problem. The problem is when it makes the same mistake twice.
Before you fix anything, look at what your AI actually did. Peek at which files it opened and what it pulled. That tells you whether the problem is the prompt (didn't ask the right thing) or the files (didn't have the right context).
The fix itself is small but consequential. Once you know what went wrong, write the correction down. Every time. Not in the chat. In the file. In Issue 1's setup that's the patterns/ folder, but anywhere structured works. Tell it: "save this so you don't do it again."
Then check the next week: did it stick? If not, sharpen the note. Make it more specific. The lesson holds when the file is good enough that your AI can read it next Tuesday and act on it without you having to repeat yourself.
This is what makes the system compound. The Monday review gets better because last week's lessons are in the file. The Friday post-mortem gets sharper because the AI has more to compare against. The job description gets tighter because the AI has learned, from a hundred small corrections, what it should and shouldn't try.
Without this loop, the article is a list of things to set up once and then watch slowly stop working. With this loop, it's a system that gets better every week.
Mistakes happen. Repeats don't have to.
The whole thing in 30 seconds
The first time I tried to set up my AI employee, I gave it a title and waited for results. Now I give it a job, a cadence, the right to push back, and the chance to learn. And it shows up to work. That's not a clever assistant. That's a chief of staff.
A real job has four parts: a scope, a cadence, the right to disagree, and the chance to learn. A title isn't enough.
The minimum useful data set for a chief-of-staff role: calendar, docs, tasks, team chat. Skip the rest.
Three standing meetings make a hire real: daily check-in, Monday review, Friday post-mortem. Daily is the engine.
Write two kinds of rules into the setup: when to speak up (disagreement) and when to act (low/medium/high risk tasks).
When it gets things wrong (and it will), write the correction into the file, not just the chat. That's the lesson that makes everything else compound.
What should I cover next? Hit reply to let me know!
p.s. I’m not a fan of ads, but Wispr Flow is worth the shout out. I use it daily and it saves me serious time. Check them out 👇
Stop re-prompting. Say it right the first time.
Voice-first prompts preserve the nuance you cut when typing. Speak once, paste into any AI tool, get results that don't need a follow-up. 89% of messages sent with zero edits.