What goes in a rules file
Developers have used one for a while: Claude Code reads a file called CLAUDE.md at the start of every session, and Anthropic's own guidance (read 22 September 2026) says to add to it when "Claude makes the same mistake a second time" or "you type the same correction or clarification into chat that you typed last session".
That is the whole method, and it works just as well for emails as for code.
- Always: the things you want in every reply. "Give the answer first, then the reasoning."
- Never, and what to do instead: "Never invent a price. Write [not stated] and ask."
- When you are not sure: ask one question, or guess and label it. You pick.
- Choices already made: spelling, date format, tone, length. Decided once, so it stops offering the alternatives.
- What done looks like: "An email is done when it is under 150 words and has one ask."
Keep it to a dozen rules or so, each one earned by a mistake you have actually had to correct. Anthropic's own advice to prompt writers (read 22 September 2026) is: "Think of Claude as a brilliant but new employee who lacks context on your norms and workflows."
Its test for any instruction is to show it to a colleague: "If they'd be confused, Claude will be too."
Where it lives in each app
- Claude: click your initials in the lower left, then Settings, then "Instructions for Claude". It applies to every chat (Anthropic's help page, read 22 September 2026). If you use Projects, rules that only apply to one kind of work go in that Project's instructions instead.
- ChatGPT: Settings, then Personalization, then Custom Instructions, with "Enable customization" switched on. Free and Go plans get 1,500 characters, and Plus, Pro, Business, Enterprise and Education plans 5,000 (OpenAI's help page, read 22 September 2026).
- Gemini: at gemini.google.com, click "Settings & help" at the bottom, then Personal Intelligence, then "Instructions for Gemini", then Add. It needs a personal Google account (Google's help page, read 22 September 2026).
- Claude Code: a file named CLAUDE.md in the folder you work in. Claude reads it at the start of every session as guidance, so it can still slip.
The prompt: write my rules file from the things I keep correcting
Replace each bracketed line with your answer, then send. It asks about anything vague, turns each correction into one rule, trims the list to what will actually be followed, writes the file, and gives you a test message to check it took.
You are writing my rules file: a short list of standing rules that I will paste into my AI's settings so it behaves the same way in every chat. I am tired of making the same corrections over and over. Your job is to turn those corrections into rules, keep the file short enough that every rule still gets followed, and refuse to invent a rule I never asked for.
WHAT I KEEP CORRECTING (I fill this in once)
- The tool I will paste this into: [e.g. "Claude, Instructions for Claude", "ChatGPT custom instructions", "a Claude Project", "Gemini, Instructions for Gemini", "Claude Code, my CLAUDE.md file"]
- What I mostly use it for: [e.g. "emails to clients and my newsletter", "planning and admin", "research for a small business", "writing code for my side project"]
- The corrections I make most often, with a real example of each: [one per line. E.g. "It adds a cheery opening line and I delete it every time: last week it started a complaint reply with 'I hope this finds you well'", "It gives me five options when I asked for one", "It uses the spelling I do not want", "It makes up a price when it does not know one"]
- Things it must never do, and why: [e.g. "never send, post or delete anything without asking, because it once drafted and 'sent' a reply I had not read", "never guess a figure"]
- Things I always want, every time: [e.g. "ask one question if the task is unclear, then start", "give the answer first and the reasoning after", "end with what you were unsure about"]
- The choices I have already made, so it stops offering the alternatives: [e.g. "spelling: [British or American]", "dates: 22 September 2026, never 9/22", "tone: plain, warm, no emojis", "length: under 150 words unless I say otherwise"]
- What "done" looks like for my usual jobs: [e.g. "an email is done when it is under 150 words, has one ask, and I could send it without editing"]
- What it should do when it is not sure: [e.g. "ask me one question", "give its best guess and label it a guess", "stop and say what is missing"]
- Rules other people's files have that I like: [paste any, or write "none"]
BEFORE YOU WRITE ANYTHING
Read my answers, then ask me up to four questions, only about things that would change a rule: a correction I gave without an example, a "never" with no reason, or two answers that contradict each other. Offer options where you can ("do you want it to ask first, or guess and label it?"). Wait for my answers. If everything is clear, say so and go on.
STEP 1: TURN EACH CORRECTION INTO ONE RULE
For every correction, write one rule in plain words, in the form the AI will read: what to do, and when. A rule earns its place only if it comes from a real mistake or a choice I have actually made. If a correction and an "always" say the same thing, keep one. If I gave you a vague one ("be better at emails"), write [NEED: what exactly went wrong last time] instead of guessing a rule.
STEP 2: KEEP ONLY THE RULES THAT WILL BE FOLLOWED
A long file gets skimmed, so aim for twelve rules or fewer, and never more than twenty. If I gave you more than that, keep the ones that fix the mistakes that cost me most time, and list the rest under "Left out, add later if it comes back". Make every rule specific enough that the AI could be caught breaking it: "keep it short" fails that test, "under 150 words unless I say otherwise" passes. Where a rule says "never", pair it with what to do instead.
STEP 3: WRITE THE FILE
Write the rules file in exactly this shape, in plain text, with no introduction:
MY RULES
Always: [the rules that apply to every reply]
Never: [each with what to do instead]
When you are not sure: [my answer from the set-up, as one rule]
Choices already made: [spelling, dates, tone, length, format, one per line]
What done looks like: [one line per usual job]
Keep it short enough to read in under a minute. Write it for the AI, in the second person ("Ask one question if the task is unclear, then start"), and put nothing in it about who I am, since that belongs in a separate note about me.
STEP 4: TELL ME WHERE IT GOES AND HOW TO TEST IT
Based on the tool I named, tell me in two or three lines where to paste it. Then give me one test message to send in a fresh chat that would break at least two of the rules if the file were not working, and say what a passing reply looks like.
RULES FOR YOU
- Never invent a rule, an example or a mistake I did not give you. Every rule must trace to one of my answers.
- Do not add generic advice ("be helpful and accurate"). The AI already does that, and it pads the file.
- Do not merge two different rules into one long sentence.
- If a rule I asked for would stop the AI doing something I also asked for, point it out and ask which wins.
CHECK THE FILE BEFORE YOU SHOW ME
1. Count the rules and show the number.
2. Every rule traces to an answer I gave. Name any that do not, and remove them.
3. Every "never" has an "instead".
4. Nothing in the file describes who I am.
5. Name the one rule you think the AI is most likely to break anyway, and why.
What a finished one looks like, for someone who writes client emails and a newsletter
MY RULES
Always:
Give me the answer first and the reasoning after.
Ask one question if the task is unclear, then start.
End with anything you were unsure about.
Never:
Open with a pleasantry. Start with the point.
Invent a price, a date or a name. Write [not stated] and ask.
Send, post or delete anything. Draft it and stop.
When you are not sure:
Give your best guess and label it a guess.
Choices already made:
American spelling.
Dates as 22 September 2026.
Plain and warm, no emojis.
Under 150 words unless I say otherwise.
What done looks like:
An email is done when it has one ask and I could send it without editing.
Why deciding in advance works
A model launched last week is built on it. Jev, released on 15 September 2026 by TypeSafe AI, only answers from a list you give it. You give it the list of answers it is allowed to choose from, and it picks one.
Its makers claim that because the possible outputs are "defined in advance", it "can't hallucinate". That is their wording, and what it means is that it cannot make something up that is off the list.
A rules file gives your chatbot a small dose of the same discipline. The choices are made before the chat starts, so the AI spends its effort on the job in front of it rather than on guessing how you like things.
The honest bit
- A chatbot can still break a rule. Jev cannot go off its list, but a chatbot reads your rules and then writes freely, so a rule gets ignored now and then. The test message the prompt gives you is how you check, and a rule that keeps failing is usually too vague.
- Test it in a fresh chat. ChatGPT applies custom instructions to chats already open as well as new ones, and Claude applies its instructions to all your conversations. A new chat is still the cleanest test, because nothing said earlier can muddy it.
- Rules pile up. Every few weeks, delete any rule for a mistake that has not come back.
Write yours tonight
Open the last five chats where you corrected something, copy the corrections into the prompt, and paste the file it gives you into your settings. Then send its test message in a fresh chat.
A few quick questions
Is this the same as an about-me file?
No. The about-me file says who you are and who you write for. The rules file says how the AI should behave. Keep them separate, because the rules stay the same when the job changes, and the context does the opposite.
Can I use one rules file in Claude, ChatGPT and Gemini?
Yes. The prompt writes a rules file that works in any of them, so paste the same file into each app's instructions box. Each app can only read its own, so you keep one master copy in a note and paste it into all of them.