It's still going to make mistakes, isn't it?

Yes. The rules file is one layer of checking, not the last one. What it changes is the rate: every time the agent goes off the rails you add a line, and the list shrinks the error rate from roughly every third time to a few percent. Brad's agent now goes off track well under 10% of the time.

I don't have time to write rules.

It's ten lines. The template is already written; you fill in the blanks. After that, each irritation adds one line, at the moment it happens. That is the whole maintenance.

Isn't it easier to just do it myself?

One operator with no technical background spent two weeks dragging the publish folder in by hand before learning the agent could deploy itself. Anything you can do through an API, the agent could do. The file is how you tell it that once instead of every time.

Which agent does this work with?

It was written for Claude Code. The file is plain text, so any coding agent that reads a file in your project can use it the same way. The five phrases are things you say, so they work anywhere you type.

Rules for Working With Me A free file from

Free · the agent operating manual

Your agent plays dumb because it assumes you know what you want.

Rules for Working With Me is the one-page file you hand your agent so it does the job instead of handing it back to you. Ten lines, five phrases, one decision rule. Fill it in tonight, paste it in tomorrow morning, and the first session after that is the one where it stops asking and starts finishing.

Every AI course teaches you prompts. Nobody tells you the agent forgets everything the moment the window closes, and fills the gap with the safest guess it has: you.

A dark desk at night with one terminal window glowing violet on the monitor and an empty chair

New here?Two minutes: the two weeks one owner lost, and the one line that ended it

How I know this is real

Two weeks of dragging a folder by hand. One line ended it.

AND

An insurance agency owner with no marketing background, a month into working with agents, high-fiving her partner every day at what it was already doing for them. “Way more than we’ve ever done.”

BUT

For two weeks she dragged the publish folder into her hosting account by hand, every single time. The agent never said it could deploy for her. It assumed she knew. “For two weeks I was dragging the publish folder in there.”

THEREFORE

One line in a file, “do it through the API, not by giving me steps to click,” and it deploys itself now. Two months in, she runs her own task board. Her agent built it.

Names left off; the words are hers, from a live call.

Why it keeps handing you the work

If longer prompts got you shorter results, that isn’t you. English is lossy.

You hired a 200-IQ employee who can’t see your screen, doesn’t know your business, and forgets everything when the window closes. So it does the cautious thing: it hands you a list of steps, or it stops and asks, or it says it read the skill when it skimmed it.

These agents don’t work autonomously unless you tell them that’s what you want. Anything you can do through an API, the agent could do. It just needs to be told, once, in a file it reads every time.

That file is the whole product. Everything below is what goes in it.

01“it had me go into the CRM, and it was manually giving me all the steps to do… and then three days later it said, ‘Oh, I can do this for you.’”

02“it will work for a little while and then it’ll stop and ask me questions, and I really want it to run longer”

03“it’ll say it referenced them, but it wasn’t really”

Three members, three calls, the same complaint. Verbatim.

The enemy

The reset.

Every window opens at zero: no memory of your business, your standards, or the last time it made you do the work. So it defaults to the safest guess, you. A chat box was never built to hold a working relationship. A file is. “Entropy is the steady state of AI.”

Why now

The good old days are this year.

The smartest models on earth are being sold for about $200 a month, a 200-IQ employee at a price that will not stay there. Write the file now and every session from here compounds on it. Skip it and you re-explain your business to a stranger every morning, at whatever the price becomes.

Who’s telling you this

Someone 2,500 hours in.

Brad Hart runs his company on agents: roughly 400 skills codified, and his own frustration down from about 40% to about 5% once this file existed. Everything on this page was mined from five Optimus Mastermind calls in the last five weeks. Every quote is verbatim. Every number is as said.

Part one · the file

Ten lines it reads before every job.

How you communicate. How you want it to communicate. This, not that. And the running list of things it did that irritated you: add a line every time it goes off the rails. Get excited, not frustrated. Every line is a mistake you never see again.

RULES-FOR-WORKING-WITH-ME.mdread on every session

# Rules for working with me
## How I communicate
- I talk in outcomes, not steps. "Get it live" means everything up to live.
- I dictate. Expect typos. Ask if it's truly ambiguous, not if it's just messy.
- When I say done, I mean landed and verified, not written.
## How I want you to communicate
- Lead with the answer. Details under it.
- Ask every question at the start, in one batch. Then run to the end.
- Report what you did with proof (the output, the URL, the commit), not what you planned.
## This, not that
- Do it through the API / MCP / CLI, not by giving me steps to click.
- Read the docs, don't guess.
- Finish the job, don't sandbag. "That could work" is not done.
## Things you did that irritated me (add a line every time)
1. Gave me manual CRM steps for three days, then said "oh, I can do this for you."
2. Said you referenced the skill. You hadn't.
3. …

Part two · the five phrases

“There’s got to be a phrase I can start with.”

There is. Five of them. They tell the agent what it’s able to do versus what you’re doing. Say them out loud, paste them in, put them at the top of the file. They come up on every call because they work.

1
Before anything you'd do by hand
“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.”
2
Before a build
“Ask me every question you have now, then don't stop until it's done. No sandbagging.”
3
When it hands you steps
“Do everything you can possibly do yourself that doesn't require me. When you truly can't, give me the exact step-by-step and the links, and I'll do it.”
4
End of a session
“Write me a handoff file; I'm opening a fresh window.”
5
Any time you hand it a skill
“Load the skill and confirm you actually read it.”

Part three · the decision rule

A hard rule is code. A judgment call is the agent.

Code

Runs the same every time.

The send window. Who gets tagged what. The price. The one thing that must never ship without a check. If you’d fire a person for getting it wrong twice, it isn’t a judgment call. Have the agent write it as code once, and it stops drifting.

Agent

Needs reading, weighing, taste.

Which lead to answer first. How to phrase the follow-up. What to cut from the draft. That is what you bought the 200-IQ employee for. Give it the rules file so it knows how you decide, then let it decide.

When it stalls, walk the chain

Agent → Skills → Tools

What does it actually need? Usually one of three things. Give it that and it goes.

A loginit can't see the dashboard you're looking at until it can sign in
An env varthe key is on your machine, not in the window it's working in
An API keythe tool exists; nobody handed it the credential

What changes when the file exists

“I’m down from like 40% frustration to like 5% frustration.”

Three outcomes from the calls this was mined from. Names left off; numbers as said.

40% → 5%

Brad’s own frustration, before and after the rules file existed. “Because I’ve gotten better at communicating.” The agent now goes off track well under 10% of the time.

2 weeks

“For two weeks I was dragging the publish folder in there.” An operator with no technical background, before learning the agent could deploy itself. She now runs her own task board, built by her agent.

Ch. 7–16

A member front-loads every question before a build and the agent “just goes to work.” It wrote chapters 7 through 16 of his book, in his voice.

“Using AI is more mentorship than anything else.” — from the same calls.

What you get

Fill in the template tonight. Paste it in tomorrow morning.

The fill-in rules file. The ten lines, with the blanks marked.
The five phrases as copy-paste blocks, each with the moment to use it.
Three short lessons: the file, the phrases, code or agent.
100+ other free resources in the same library, one sign-in link, no password.

Price: your email. Unsubscribe anytime.

What you have to lose

A file your agent reads before every job is what turns a 200-IQ guesser into someone who does the work.

Option one

Keep re-explaining your business to a blank window every morning, and keep doing the parts it hands back.

Option two

Write your own rules from scratch, one irritation at a time, over the next few months. It works. It is slow.

Option three

Take the file tonight. Paste it in tomorrow. The people who act on this are the ones whose agents actually work for them.

Fair questions

The objections I hear every time I show this file.

It's still going to make mistakes, isn't it?: Yes. The rules file is one layer of checking, not the last one. What it changes is the rate: every time the agent goes off the rails you add a line, and the list shrinks the error rate from roughly every third time to a few percent. Brad's agent now goes off track well under 10% of the time.
I don't have time to write rules.: It's ten lines. The template is already written; you fill in the blanks. After that, each irritation adds one line, at the moment it happens. That is the whole maintenance.
Isn't it easier to just do it myself?: One operator with no technical background spent two weeks dragging the publish folder in by hand before learning the agent could deploy itself. Anything you can do through an API, the agent could do. The file is how you tell it that once instead of every time.
Which agent does this work with?: It was written for Claude Code. The file is plain text, so any coding agent that reads a file in your project can use it the same way. The five phrases are things you say, so they work anywhere you type.