What Are Rules For Working With Me (For an AI Employee)?

Rules For Working With Me is a one-page file you write once and your AI employee reads before every shift. It covers four things: how you communicate, how you want the agent to communicate back, a this-not-that list, and a running list of every time the agent irritated you. It is not a prompt and it is not a job description. It is the standing operating rules a new hire would absorb in their first month, written down so the agent absorbs them in the first second.

Why does an AI employee need a rules file at all?

Because English is lossy. When you tell a human employee "handle the Johnson renewal," they fill in the gaps from a year of watching you work. An AI employee has no year of watching you. It fills the gaps with the safest guess it can make, and the safest guess is almost always to hand the work back to you with a question or a half-finished draft.

That is the pattern that came up on five straight Optimus Mastermind calls between August 26 and September 23, 2026. Different businesses, different agents, same complaint. One member said it plainly: "I think that's the biggest learning curve for me is the communication with the agent." The agents were capable. The owners were capable. The handoff between them was broken, and nobody had written down the rules of the handoff.

The rules file fixes that. It is the written version of the context a human picks up by osmosis. The agent reads it, every time, before it touches your work.

What is actually in the file?

Four short sections. The whole thing fits on one page, and it should stay on one page.

How I communicate. Are you terse? Do you talk in fragments at 6am from your phone? Do you say "do the thing" and expect the agent to know which thing from the last three messages? Tell it. An agent that knows you write in shorthand stops asking you to clarify shorthand.

How I want you to communicate. Lead with the answer or lead with the reasoning? Bullets or prose? Do you want a status line at the end of every task, or only when something went wrong? Most owners have strong opinions here and never state them, then get annoyed at every reply.

This, not that. Concrete substitutions. "Ask every question up front, not one at a time." "Use the API, not a manual walkthrough for me to follow." "Finish the job, not a plan for the job." Each line removes one recurring argument.

The irritation list. This is the section that does the heavy lifting. Every time the agent goes off the rails, you add one line describing what it did and what it should have done instead. It is the cheapest performance review ever conducted. On my own agents, frustration dropped from roughly 40 percent of interactions to about 5 percent once this list existed, and the agent now goes off track well under 10 percent of the time.

What goes with the rules file?

The file is the core, but two small companions make it work in practice.

The first is a set of five phrases that tell the agent what it can do versus what the human is doing. Examples: "Is there an API or MCP that lets you do this for me instead of me doing it by hand? Don't guess, read the docs." Or: "Ask me every question you have now, then don't stop until it's done. No sandbagging." These are the sentences that end the back-and-forth. We cover them in detail in how to write the rules in thirty minutes.

The second is a decision rule: a hard-and-fast rule is code, a judgment call is the agent. If something must happen the same way every time, do not ask an agent to remember it; make it a script, a cron job, a validation step. If something needs judgment, that is what you hired the agent for. Blurring those two is how owners end up babysitting work that should have been automated and automating work that needed a brain.

How is this different from a prompt or a system prompt?

A prompt is an instruction for one task. The rules file is the standing context under every task. You will write hundreds of prompts this year. You will write one rules file and edit it a line at a time. The full breakdown is in AI employee rules vs prompts, but the short version: prompts tell the agent what to do today, rules tell it how you want things done, period.

It is also different from a job description. A job description says what the role is responsible for. The rules file says how to work with you specifically. A great hire with the wrong rules file still drives you crazy. A decent agent with a tight rules file quietly does the work.

Who is this for?

Business owners, mostly in the $5M to $50M range, who are about to put an AI employee on the payroll or already did and are frustrated with it. You do not need to be technical. One operator on the calls had no technical background at all and spent two weeks dragging build folders by hand before she learned the agent could deploy itself. Her rules file now has a line about that, and she runs her own task board that her agent built.

If you are not sure you need any of this yet, read do I even need rules for my AI employee first. If you already know you do, the template is free and takes about thirty minutes to fill in.

FAQ

Is Rules For Working With Me a prompt?

No. A prompt is a one-time instruction for a specific task. Rules For Working With Me is a standing file the agent reads before every task. It describes how you communicate, how you want the agent to communicate, your this-not-that preferences, and a running irritation list. You write prompts constantly; you write this file once and edit it a line at a time.

How long should the rules file be?

One page. If it grows past that, you are putting task instructions or job descriptions in it, and those belong elsewhere. The irritation list is the only section that grows, and it grows one line at a time, only when the agent actually did something wrong.

Does it work with any AI agent?

Yes. The file is plain text. It was built around coding agents like Claude Code because that is what the Optimus Mastermind members were using, but any agent that can read a file or accept a system instruction can read it. The content is about you, not about the tool.

What is the irritation list?

A running list, one line per incident, of every time the agent did something that annoyed you and what it should have done instead. The agent reads it before every shift. It is the section that produces the biggest drop in error rate, because it is built from your real failures rather than generic advice.