My AI turned a passing comment into a rule.

I use Claude Code and Codex most days. Recently I asked Claude whether it was worth trying DeepSeek to bring my costs down.

It gave me two reasons not to. The first was fair enough: DeepSeek's own API stores data in China, and a client asking where their data goes would not love that answer. The second was that switching would “reverse a decision you've already made.”

I had to go and look.

The decision turned out to be one line in my project notes: models are OpenAI and Anthropic only. Claude had written it down while we were planning something else, and I'd said yes without much thought. I never meant it as a rule. Once it was in the file, though, Claude treated it as one.

Codex does the same thing. So do ordinary chat apps. Mention something once and it can keep coming back with far more weight than you gave it.

The brief is all the agent has.

Every agent I set up reads a short brief before it starts a job. It's usually a file called CLAUDE.md or AGENTS.md. It covers what the business does and how it talks to customers, along with anything the agent must never do.

The agent has no way to tell a passing remark from a firm policy. Both are sentences in the same file.

So writing a brief is a balancing act. Leave something out and the agent fills the gap with a guess. Write something loosely and it may hold on much harder than you meant. Here's what I do about that now.

Say how firm it is.

Fixing my model line took a few seconds:

The facts are the same. The weight is different. A colleague would pick that up from your tone of voice, but an agent only gets the words, so “for now” and “always” have to do that job.

Give the reason.

“Never quote prices over email” is a rule the agent can follow or break, and that's about it. Add “because supplier costs change every week” and it has something to reason with. It can explain the rule to a customer. It can also flag the day the reason stops being true.

Cut what doesn't matter.

Every line in a brief competes for the agent's attention. When I stuff in every passing thought, small things start getting treated as instructions and the important parts get lost among them.

My test is whether I'd say it to a new staff member on their first day. If I wouldn't, it comes out.

Update it the day things change.

People forget old decisions. Briefs don't. A line written in March keeps its full weight until somebody deletes it, so when something changes in the business, I change the brief straight away. I fixed my model line before I'd finished the first draft of this post.

Then you never have to say it again.

That's the good part. Once your tone and your pricing rules are written down properly, the agent reads them before every job and you stop repeating yourself.

It's also a big part of setting up an agent for a business. Connecting Slack to a CRM is the easy bit. The slow work is learning what actually matters to the owner, and how firmly they mean it.

If you've used an AI tool that kept getting your business slightly wrong, book a free 15-minute call and I'll show you how I'd brief yours.