Guides · AI at work
ChatGPT Custom Instructions Examples: Make Each Fix Only Once
By Josh Croom · October 27, 2026 · 12 min read
Custom instructions are a standing note ChatGPT reads before every new chat, so you stop making the same fixes over and over. The best ones are short, give a reason for every rule, and come from fixes you've already made more than once.
The third rewrite of the same email
Tuesday, 4:10 PM. You ask ChatGPT for a quick follow-up to a customer who got an estimate last week. It opens with “I hope this email finds you well,” runs four paragraphs and closes with three exclamation points. So you type “shorter,” then “less stiff,” then “don't open like that.” The third try is good, and you send it.
The next morning you open a new chat for a different customer, and the first draft makes every one of those mistakes again. Your fixes stayed behind in yesterday's conversation. Custom instructions carry those fixes into every new chat, so you make each one only once.
Why fixing it mid-chat doesn't stick
A 2025 study by researchers at Microsoft and Salesforce gave AI models the same tasks two ways: all at once in a single message, or one piece at a time over a chat.
- 39%worse on average when the task came in bit by bitacross six kinds of tasks
- 200,000+simulated chats in the studyrun by the researchers
- 15AI models testedlarge language models, tested the same way
What the study means for your chats
The models jumped to guesses early, gave a final answer before they had all the details and seldom recovered from a bad start. The researchers suggest starting a new chat with all the details up front, since that often works far better than rescuing one that went off course. A standing note gives every new chat those details before you type a word.
Where custom instructions live
In ChatGPT, custom instructions sit in Settings, under Personalization. The menu names move around from time to time, so if you can't find it, search the settings for “custom instructions.” Whatever you save there carries into each new chat you start.
Other AI tools have the same idea under a different name, often preferences or instructions, and these examples work in any of them. Some tools give you a small box, so keep a version that fits on half a page.
How to write yours, in eight steps
Most of the work happens before you open the settings, because the note is only as good as the fix list behind it.
Step 1
Keep a fix list for two weeks
Each time you tell AI it's wrong, too long or too stiff, write one line on a notepad saying what it did and what you wanted instead.
Step 2
Give each fix a reason
“Shorter” turns into “Keep emails short, because I read them on my phone.” A reason lets ChatGPT handle cases your note doesn't name, because it knows what you're after.
Step 3
Keep only the fixes that came back
If you made a fix more than once, it goes in the note. A fix you made one time can wait, because a long note buries the rules that matter most.
Step 4
Fill in the four-line card
Write what you use AI for, how it should sound, what it should always do and what it should never do, and give every line its reason.
Step 5
Let AI clean up the wording
Paste your card into a chat and ask it to clean up the wording without adding any rule you didn't write, so your words and reasons come back intact.
Step 6
Paste it where ChatGPT reads it first
Put the note in custom instructions rather than in a chat, so it's there before your first word every time, and save the half-page version if your box is small.
Step 7
Test it in a new chat
Open a new chat and ask for the thing you used to fix the most. If it still gets it wrong, change the note rather than arguing with the chat, because the note is what every future chat reads.
Step 8
Update it, and give each person their own
Add a line when a fix comes back, and cut lines that no longer matter every few months. Each person should write their own note, while facts about your business belong in one shared document that everyone pastes in.
A rule without a reason, and the same rule with one
A reason lets ChatGPT apply a rule on its own, even in a case your note never mentioned.
Without a reason
- Be shorter.
- Less formal.
- Don't make things up.
- Use bullet points.
With a reason
- Keep emails under 120 words, because I read them on my phone.
- Write the way I'd talk to a regular customer, because most of ours know me by my first name.
- If you don't know a fact, say so and ask me, because I'd rather fill a gap than fix a wrong number.
- Use a list only for steps someone will follow, because everything else reads better as sentences.
The four-line card
Copy these four lines into a note and fill them in from your fix list, ending each line with its reason.
- I use AI for: ____.
- It should sound like: ____.
- Always do this, and here's why: ____.
- Never do this, and here's why: ____.
Example 1: A home services owner
- I use AI for: replies to customer emails, follow-ups on estimates and answers to Google reviews.
- It should sound like: me talking to a neighbor over the fence, friendly and plain, with no sales talk.
- Always do this, and here's why: keep emails under 120 words and end with one clear next step, because customers read them on their phones too.
- Never do this, and here's why: never promise an arrival time or a price, because only the office can confirm those, and a wrong promise costs us the job.
Example 2: An office manager
- I use AI for: meeting notes, schedules, short memos to staff and formulas for our spreadsheets.
- It should sound like: a calm coworker who gets to the point.
- Always do this, and here's why: put the answer or the action in the first line and the details after it, because most of our staff only read the top of a message.
- Never do this, and here's why: never invent a date, a name or a policy, and ask me instead, because a wrong date on a schedule makes more work than it saves.
Example 3: A nonprofit director
- I use AI for: board updates, donor thank-you notes and first drafts of grant reports.
- It should sound like: warm and direct, the way I'd talk to a longtime volunteer.
- Always do this, and here's why: lead with what changed for the people we serve, then give the numbers. Mark any number I didn't give you with (check), because I report only figures I can back up.
- Never do this, and here's why: never use the names of people we serve or any detail that could identify them, because their privacy comes before a good story.
Example 4: A church office administrator
Our guide to AI for churches covers the rest of a church office's week.
- I use AI for: the bulletin, the weekly email, serving reminders and event announcements.
- It should sound like: a friendly member welcoming a guest, warm and simple, with no insider words.
- Always do this, and here's why: copy service times, addresses and dates exactly as they appear in my notes, because a wrong time sends a guest to an empty building.
- Never do this, and here's why: never add a prayer request, a Bible verse or a quote unless I gave it to you word for word, because we check every verse before anything goes out.
Example 5: An online store owner
- I use AI for: product descriptions, replies to customer questions and email subject lines.
- It should sound like: a shop owner who knows every product by hand, specific and a little playful.
- Always do this, and here's why: start a product description with what the customer gets and who it's for, because shoppers read the first line and decide whether to keep going.
- Never do this, and here's why: never claim a product cures or guarantees anything, and never invent a size, a material or a shipping time, because a wrong claim brings returns and complaints.
Example 6: A salesperson
- I use AI for: follow-up emails after calls, call notes and short proposals.
- It should sound like: me on a good day, when I'm curious and plain and never pushy.
- Always do this, and here's why: open with something the person said on the call and keep it to one ask per email, because it shows I listened and one ask is easier to say yes to.
- Never do this, and here's why: never use pressure lines like “limited time” or “act now,” because they cost trust with the kind of customer we want.
Example 7: How you want answers laid out
- I use AI for: research, explanations and decisions where I need to understand something new.
- It should sound like: a patient expert explaining to a smart friend outside the field.
- Always do this, and here's why: give the answer first in two sentences, then the reasons, and tell me when you're unsure, because I'd rather know what's uncertain than get a confident guess.
- Never do this, and here's why: never repeat my question back or tack on a closing offer to help more, because I read on a small screen and every extra line slows me down.
Example 8: A never list for writing
Add these lines to any other example, since they cover the habits that make writing sound machine-made.
- I use AI for: anything that someone else will read.
- It should sound like: a person talking, not a press release.
- Always do this, and here's why: use contractions and short, common words, and write the way I'd say it out loud, because people trust writing that sounds like a person.
- Never do this, and here's why: never use long dashes or open with “I hope this finds you well,” because those habits give a draft away.
The same request, after the note
Once Example 1 is saved, the same follow-up comes back short and plain, with one next step. A brand-new customer who calls or fills out your form needs an even faster reply, and our guide to speed to lead covers that first callback.
Write a follow-up to Dana about the water heater estimate we sent last Tuesday.
Hi Dana, I'm checking in on the water heater estimate we sent Tuesday. If it looks right, reply here and our office will call you to pick a time. Chris
Why ChatGPT ignores your custom instructions
When a note doesn't seem to work, check four things. ChatGPT follows a rule with no reason word for word, and it can miss the point. A long note buries the rules that matter. Two rules can fight each other, like “be thorough” and “keep it short,” and then ChatGPT has to pick one. And a long chat that already went off course still carries its old history, even though ChatGPT now applies a saved change to chats under way.
Fix the note, then test it in a new chat, because the models in the study seldom recovered from a bad start.
Is your note doing its job?
Answer five quick questions. Your answers stay on your device.
Did every rule in your note come from a fix you made more than once?
Does every rule have a reason after it?
Does the whole note fit on half a page?
Did a new chat get your most common fix right the first time?
Have you kept names, health details and money details out of the note?
Your note is doing its job. Add a line each time a fix comes back, and trim the note every few months so it stays short.
You're close. Fix the first question you answered No, then test the note again in a new chat.
Start a fix list today, because two weeks of notes will give you everything you need to fill in the four-line card.
Set it up this week
Tick each one off as you go. Your ticks stay on this device.
0 of 7 done
Talk it through with Miguel
Our in-house AI helps us with research, drafts and tracking checks, and a person checks every setup before it goes live. If you'd like to talk about how your Google Ads bring in calls and forms, and where the busywork around them could get lighter, book a call with Miguel.
Questions people ask
What are ChatGPT custom instructions?
They're a standing note in ChatGPT's settings that it reads before every new chat. You use them to say what you do, how you want it to sound and what it should always or never do, so you stop repeating the same fixes.
What should I put in ChatGPT custom instructions?
Put in the fixes you've made more than once, each with a reason. Four lines cover most needs: what you use AI for, how it should sound, what it should always do and what it should never do.
How long should custom instructions be?
Keep them short enough to fit on half a page. Some tools give you a small box, and a short note keeps the rules that matter from getting buried.
Do custom instructions change chats I already started?
Yes. Since November 2025, ChatGPT applies a saved change to your personalization settings right away, in chats already under way as well as new ones. Still test an edited note in a new chat, because a chat that went off course carries its own history.
Why does ChatGPT ignore my custom instructions?
Check four things: each rule needs a reason, the note needs to be short, no two rules should fight, and the test has to happen in a new chat. Then fix the note rather than arguing with the chat.
See how we run Google Ads for your business
Read next
- Call Log Template: What to Write Down From Every CallA ready-to-copy call log, the one question to ask every caller, and how to sort wrong numbers.
- ChatGPT Prompt Templates: Write the Weekly Ask Once and Save ItCopy-ready ChatGPT prompt templates for the jobs you run every week, and how to save each one.
- Is It Safe to Connect ChatGPT to Gmail? Set These Limits FirstWhat ChatGPT can do once it's in your Gmail, and the limits to set before you connect.