Rules vs prompts: what to write once and what to say every time
Some instructions belong in the message. Some belong in a rule your assistant reads every time. A simple test for telling them apart, with a before and after.
Most people who use an AI assistant every day have a small ritual they no longer notice. Open a chat, paste the same three sentences about tone, paste the same reminder about currency, then type the request. Some keep those sentences in a note on their desktop. Some retype them from memory and forget one about half the time.
Those sentences are rules. They are sitting in the wrong place.
Two places to put an instruction
Every instruction you give an assistant lands in one of two places.
The message is what you type in the chat box. It applies to this reply and, loosely, to the rest of this chat. Tomorrow, in a new chat, it is gone.
A rule is a standing instruction the assistant reads before every conversation. You write it once, in your assistant's settings or in a project, and it is on the desk before you type a word. Claude calls these Instructions for Claude. ChatGPT calls them Custom instructions. Cursor, Claude Code, and Codex read them from a settings page or a plain text file such as AGENTS.md.
The model does not learn from your corrections. Fixing something in a chat fixes that chat. If a correction should hold tomorrow, it has to live somewhere the assistant reads every time, and that place is a rule.
The test
For any instruction, ask one question: would I want this in a chat about something unrelated?
"Use Canadian spelling." Yes, in almost every chat. That is a rule.
"Never promise a delivery date in writing." Yes, every time you draft anything a customer will see. Rule.
"The Henderson proposal is due Friday." No. Next week it is stale, and every unrelated chat would start with a note about Henderson. That goes in the message, or at most in a project for Henderson.
"This reply should decline the meeting and offer two times next week." No. That is the finish line for one task. Message.
A second test catches the rest: have I typed this twice in separate chats? If yes, promote it. The second time you correct the same thing is the moment to write it down.
A worked example
Picture a bookkeeper who types some version of this request several times a week.
Before (typed every time):
I run a two-person bookkeeping firm in Vancouver. My clients are small businesses, mostly trades. Use Canadian spelling and CAD. Keep it short and plain, no jargon, my clients are not accountants. Don't give tax advice as fact; if something depends on their situation, say so. No exclamation marks.
Write a reminder to Dave's Plumbing that their Q3 receipts are due by October 15, and that we can't file their GST return without them.
Ninety words, of which twenty-five are about Dave. The other sixty-five are the same every time, and on the days she forgot the line about tax advice, the drafts drifted into confident statements about what Dave could deduct.
After, the rule (written once, in account settings):
About me- I run a two-person bookkeeping firm in Vancouver, BC. Clients are small businesses, mostly trades.- Use Canadian spelling. Money is in CAD.How to answer- Plain language. My clients are not accountants; explain any term they might not know.- Keep client messages under 120 words unless I ask for more.- No exclamation marks.When you are unsure- Never state tax treatment as fact. If it depends on the client's situation, say so and suggest they confirm with us.
After, the message (typed today):
Reminder to Dave's Plumbing: Q3 receipts due October 15, we cannot file their GST return without them.
The message dropped from ninety words to eighteen. More important, the tax line is now on the desk every time, including the days she is in a hurry, which were the days it used to go missing.
What does not belong in a rule
The failure on the other side is just as common. Someone discovers rules and pours everything in: every client, every project, a full style guide, a list of forty preferences. Two things go wrong.
First, the rules start to fight. "Keep it short" and "always include full background" cannot both win, and when they conflict the assistant picks one without telling you which.
Second, attention thins. A rule buried in line thirty-seven of a long block carries less weight than a short list the assistant can hold in view. Around fifteen lines is plenty for account rules. Past thirty, some of it belongs somewhere narrower.
The somewhere narrower has two options:
If the instruction is about...
Put it in
One client, product, or ongoing piece of work
Project instructions
How to do one recurring task, step by step
A skill
Projects hold background for one body of work, plus the files the assistant should work from. Skills hold a method: the steps, the output format, the checks. Your first skill in fifteen minutes walks through building one.
How to word a rule so it gets followed
The rules that work are the ones the assistant can check itself against. "Be concise" is a wish. "Keep answers under 200 words unless I ask for more" is a rule, because anyone can look at an answer and say whether it was followed.
Three habits help:
Say what to do, not only what to avoid. "Don't be salesy" leaves a gap. "Describe what the service does and who it is for, then stop" fills it.
Give the reason when it is not obvious. "Never promise a delivery date, because turnaround depends on the client sending files" lets the assistant apply the rule to a text message you never listed.
Test one change at a time. Edit a rule, open a fresh chat, ask for something the rule should affect. If nothing changed, the wording is too vague or the rule is in the wrong layer.
The short version
Say the specifics of today's task in the message. Write down everything you would otherwise repeat. Keep the written part short enough that you would notice if a line went stale.
Start here
The Rules Builder is the fastest way to a first rules file: eight plain questions, and you get text ready to paste into Claude, ChatGPT, Cursor, Claude Code, or Codex. It runs in your browser and your answers stay there.
For the full method (the three layers, a starter set to edit, project instructions, and keeping rules alive as your work changes), the AI Field Guide is $39 USD, one-time. If you would rather start from filled-in files than a blank page, the Skills & Rules Starter Kit is $149 USD.
The AI Field Guide teaches the whole method in plain language: how assistants read you, how to ask, how to check the answer, and how to write rules and skills. For Claude, ChatGPT, Cursor, and Codex.
No charge until checkout opens.
Damon Blais
Operator, Albino Geek Services
Works with AI assistants and coding agents every day, and writes down what makes them do what you mean. Plain language for everyday users, the deep end for builders.
