Teach AI Once: The Brief Once Method for Rules

You tell AI to stop using a certain phrase. It stops. Next week it uses the phrase again. Feedback disappears, and rules stay. That is the whole idea of this guide.

The brief once method turns every correction into a written rule your AI loads each time. You say it one time and it holds.

After any correction, say one sentence: "Make that a rule." Then have the AI save it in your memory file, so it applies to every future job. Anthropic's docs list a repeated correction as a signal to add a line to CLAUDE.md.

Why corrections do not stick

When you correct AI inside a chat, the correction lives only in that chat. A new session has no record of it. So you repeat yourself, and you start to think the tool is not learning.

It is not that it cannot follow a rule. It never had the rule. Move the correction out of the chat and into a file it reads every time, and the problem should shrink. The docs are clear that these files are context, not enforced rules, so a saved rule makes following it more likely, not certain. For a rule that must never be broken, such as blocking an action, the docs point to hooks instead, so ask a technical helper to set one up.

The method in four steps

1. Give the feedbackSay exactly what was wrong and what you wanted. "Never leave a single word alone on its last line."
2. Say whyOne short reason helps it apply the rule well. "It looks unfinished on a phone."
3. Make it a ruleSay: "Make that a rule and save it in the memory file."
4. Check it savedOpen the file and read the new line. If it is vague, ask for a sharper version.

The Halftime lesson on teaching AI your taste uses this sentence after reviewing a full week of posts: apply the feedback to every post, then save each point as a rule so the mistake does not return.

Make rules specific enough to work

A rule has to be something a person could check. Compare these.

If you cannot tell whether a rule was followed, the AI cannot either. Rewrite until a stranger could check it.

Keep a short list, not a long one

Short, clear rules beat long, fuzzy lists. Anthropic's docs say longer files reduce how well instructions are followed, and that contradictory instructions may be resolved arbitrarily. Group them by what they cover, such as writing, design, and safety. When two rules clash, delete one. Review the list each month and remove anything that no longer applies.

Rules do their best work next to the context in your memory file. The memory says what your business is. The rules say how to behave inside it.

Rules for safety, not just style

Style rules save time. Safety rules save trouble. Good ones to write early:

• Draft first, never publish or send without my approval.
• Never delete files. Move them to an archive folder instead.
• Never use a number or claim I have not given you.
• Ask me before anything that costs money.

Say what you like, too

Rules made only from complaints produce nervous, bland work. When something comes out right, say so and save that as well. "Make that a rule: the first line of each email names the benefit." Good examples teach as well as bad ones do.

Keep your own voice in it

Rules help the AI avoid your pet hates, but they do not give it your heart. You still have to review the output. Work posted without your feedback tends to sound like everyone else's, and readers notice.

A weekly habit

Once a week, review a batch of what the AI produced in one sitting. Write every note down, turn each into a rule, and let it apply them across the whole batch. It takes a few minutes and the quality of week two beats week one.

A worked example: twelve rules for a small shop

Say you sell handmade candles online and have just reviewed a week of product descriptions the AI wrote. Your notes become rules like these (an illustrative set):

• Describe the scent in three words or fewer before anything else.
• Never claim a burn time unless I have given it to you.
• No exclamation marks.
• Always end with the care instruction line.
• Use "candle", never "luxury candle".
• Prices are written with the currency symbol first.
• One short paragraph, then a list of three details.
• Never mention a competitor.
• Draft only, I publish.
• Ask me when a fact is missing.
• Keep product names exactly as spelled in the product file.
• Do not use the words on my banned list.

This is an illustration of the shape of a rule set, not a list from a real business. Notice that every line can be checked by reading the output.

What to do when a rule is ignored

Ask the AI to quote the rules it has loaded. If the rule is not there, it was never saved, so save it. If it is there and was ignored, make it sharper, move it near the top, or add a reason. If two rules clash, it will pick one, so remove the clash.

Where rules end and judgement begins

Rules cover what you can state. Taste cannot always be stated. For that, keep a folder of examples you love and point the AI at it. Then review its work yourself and turn any new pattern into one more rule.

Questions people ask

What is the brief once method?

It is turning each correction into a saved rule, so you give an instruction one time and the AI follows it every session after.

Where do the rules get saved?

In a CLAUDE.md file in your project folder, which Claude Code loads at the start of each session. Claude Code also has auto memory, where it writes its own notes from your corrections, so check what has been saved.

How many rules should I have?

As few as work. Anthropic suggests keeping each file short, under 200 lines. Remove any that clash or go stale.

Will AI always follow the rules?

Not perfectly. Check the output, and sharpen any rule that was ignored or misread.

Next step

See the whole habit on camera in Lesson 1.1, and in Teach AI your taste for a full week review.

Go deeper with the lesson

This guide gives you the idea. Lesson 1.1, Set up AI to do real work for you takes you through it step by step, in plain words, so you can do it for your own business.