What AI Employee Rules Should Cover (The 7-Section Checklist)

AI employee rules should cover seven things: the business context the agent is working inside, how you communicate, how you want it to communicate back, a this-not-that list, what it is allowed to do without asking, what access it needs to do the job, and a running irritation list. The core four (how I communicate, how I want you to communicate, this-not-that, irritation list) are the Rules For Working With Me template. The other three are the surrounding scaffolding that stop the agent from stalling for reasons that have nothing to do with communication. All seven should still fit on one page.

1. Who you are and what the business does

Three to five sentences. What the company sells, who buys it, roughly what size, and what you personally own in the business. An AI employee working for a $20M industrial distributor should not write like it works for a boutique agency, and it will unless you tell it. Include the one or two things that make your business unusual, because those are exactly the things a generic agent gets wrong.

Keep it short. This is orientation, not a company history. If the agent needs to know your full product catalog, that is a reference document it can be pointed at, not a rules section.

2. How I communicate

How you actually write and talk, not how you would like to. Fragments or full sentences. Phone or desktop. Whether a string of messages is one instruction or several. Whether "fix it" means the last thing you mentioned or the whole area. What you hate being asked twice.

This is the section that stops the agent from treating your shorthand as ambiguity. English is lossy; the agent fills gaps with the safest guess, which is handing the work back to you. If it knows your shorthand, there are fewer gaps to fill.

3. How I want you to communicate

What you want back. Answer first or reasoning first. Bullets or prose. A one-line status at the end of each task: done, blocked, or needs you. What counts as worth interrupting you for. Whether you want to see the work in progress or only the finished result.

Two lines that belong in almost every version of this section: "When you are blocked, tell me exactly what you need: a login, an env var, an API key." And: "When you truly cannot do something, give me the exact step-by-step and the links, and I will do it." Those two lines alone remove most of the vague "you'll need to set that up" replies that waste an afternoon.

4. This, not that

Concrete substitutions, one per line. The five that came up on every Optimus Mastermind call:

• Use the API or MCP, not a manual process for me to do by hand. Read the docs, do not guess.
• Ask every question now, then do not stop until it is done. No sandbagging.
• Do everything you can yourself; only hand me what truly needs me.
• Write a handoff file before I open a fresh window.
• Confirm you actually read the skill or document I told you to load.

Add your own as they come up. This section and the irritation list are where the file earns its keep; the thirty-minute writing guide walks through both.

5. What you can do without asking

Where the line sits between act and check in. Most owners never state this and then are surprised when the agent either asks permission for everything or does something irreversible without a word. Write down the categories: what it can do freely, what it does and then tells you, and what it stops and asks about first. Spending money, sending anything to a customer, deleting anything, and touching production systems are the usual stop-and-ask items. The full treatment is in when an AI employee should escalate to a human.

Apply the decision rule here: a hard-and-fast rule is code, a judgment call is the agent. If something must never happen, do not rely on a rule; put a technical block in place. Rules are for the judgment zone.

6. What you need to do the job

Access. This is the section that nobody thinks belongs in a rules file and that prevents more stalls than any other. The checklist we teach is agent, then skills, then tools: when the agent stalls, ask what it actually needs. Usually the answer is a login, an environment variable, or an API key it does not have.

Write down what the agent has access to and where to find it. Write down what it does not have yet and who to ask. Write down the standing instruction: if you discover you need access you do not have, say so in one line with the exact name of the thing, and stop. An agent that knows how to ask for a key gets the key in a minute. An agent that improvises around a missing key costs you a day.

7. The irritation list

A heading, empty on day one. Every time the agent does something that irritates you, add one line: what it did, what it should have done. It reads the list before every shift. On my own agents this is the section that took the error rate from roughly every third task to a few percent. It works because it is built from your actual failures, not from a generic best-practices list someone else wrote.

What should not be in the rules

Task instructions. Project-specific context. Your full product catalog. Anything that applies to one job and not the next. All of that belongs in prompts or reference documents the agent can be pointed to. A rules file that grows past one page stops being read with care, and the whole point is that it gets read, carefully, every single time. If yours is already long, the rule mistakes guide explains what to cut.

FAQ

Do all seven sections need to be filled in on day one?

No. Sections two through four take about thirty minutes. Section one is a few sentences. Sections five and six can start as three or four lines each. Section seven, the irritation list, is deliberately empty on day one and fills itself as the agent makes real mistakes.

Why is access part of the rules file?

Because a large share of agent stalls are not communication problems; they are a missing login, env var, or API key. Writing down what the agent has, what it lacks, and how to ask for more turns a day of improvising into a one-line request you can answer in a minute.

Can the rules file be longer than one page if I have seven sections?

It should not be. Each section is a handful of lines. If it is running long, you are putting task instructions or reference material in it. Move those to prompts or separate documents the agent can be pointed to, and keep the rules file short enough to be read carefully every time.

Which section matters most?

The irritation list. It is built from your agent's actual mistakes rather than generic advice, and the agent reads it before every shift, so each mistake is made once. The this-not-that list is a close second because it removes the recurring arguments before they start.