ChatGPT custom instructions for a business should hold four things: who you are and who you write for, the tone you want with two or three examples, the format rules for each common job, and a firm instruction to ask rather than invent prices, dates or policies. Keep the set under about 1,500 characters and test it.
The Complete Excel Course with AI Integration
From formulas and pivot tables to Power Query, DAX, VBA and AI with Copilot, ChatGPT and Claude — hands-on from the first lesson.
By Sagnik Bhattacharya and Virginia Thorn, Primetime Emmy Award winner
Buy nowAnything specific to one job belongs in a project, not in your personal set. OpenAI says project instructions override your global custom instructions inside the project, so a short general personal set plus a few job-specific projects is the setup that holds up. Below are three filled-in sets, a list of what to leave out, and a five-prompt test.
Where do ChatGPT instructions live, and which one wins?
Project instructions win inside their project, overriding your personal custom instructions, so keep the personal set short and general.
There are three places your guidance can sit, and the order of priority matters.
| Level | Where you set it | Applies to | Who sees it |
|---|---|---|---|
| Personal custom instructions | Settings, then Personalization, then the Custom Instructions field (on mobile: Settings, then Customize ChatGPT) | All your chats, including existing ones, straight away | You only; not visible to shared-link viewers |
| Project instructions | In a project, the more-options menu, then Project settings | Only chats inside that project; override your personal set | Collaborators, if you share the project |
| Workspace-wide | Not documented for Business in OpenAI's help pages at the time of writing | Not established | Check Workspace settings in your account |
Two more facts from OpenAI's help are useful. Changes apply immediately across all chats, and there is an "Enable customization" toggle you can switch off to test without your instructions. Also, personalised temporary chats can use custom instructions, while unpersonalised ones do not use them. That last point gives you a free baseline for testing, covered below.
Custom GPTs used to be the other way to hold instructions, but OpenAI's retirement FAQ says they are scheduled to retire on 11 December 2026, with migration to plugins, where a GPT's instructions become a skill. For a team that needs reusable instructions, projects are the stable place to put them now. The tutorial on customising ChatGPT with projects, instructions and files shows the project side in detail.
What goes in a good set of custom instructions?
A good set has five blocks: who you are, who you write for, your voice, format rules for recurring jobs and a safety rule. Write them in plain sentences, not keywords.
- Who I am. What the business does, its size, who it serves. Two sentences.
- Who I write for. The typical reader and what they care about.
- Voice. Three adjectives and one banned habit, such as "no exclamation marks".
- Format rules for recurring jobs. Emails under 120 words, bullet summaries first, prices always in dollars.
- The safety rule. "If you do not have a price, date, name or policy, say so and ask. Never guess."
That fifth block does more for accuracy than any other line. A model that has been told to ask will often ask; a model that has not will fill a gap with something plausible.
A character budget
The 5,000 character limit on paid plans is a ceiling, not a target. An illustrative budget for a 1,500-character set:
| Block | Approximate characters |
|---|---|
| Who I am | 250 |
| Who I write for | 200 |
| Voice | 300 |
| Format rules | 500 |
| Safety rule | 250 |
What do custom instructions look like for a coworking space?
A coworking community manager can use one personal set that states the size of the space, the reader, the voice and an answer-first format, as in the filled-in set below.
The job: replying to enquiries, writing member newsletters and short notices. The reader is a freelancer or small-team founder deciding on a desk. Here is a filled-in personal set, illustrative, with placeholders for the details only you know.
I manage a coworking space with about 60 members, mostly freelancers and small teams.
I write enquiry replies, member notices and a monthly newsletter.
Readers: busy people choosing a desk or checking house rules. They want the answer first.
Voice: friendly, plain, practical. No exclamation marks, no corporate phrases like
"we are delighted to inform you". Use British spelling.
Format: replies under 120 words. Lead with the answer, then one next step.
Notices: a one-line subject, then three short lines at most.
Newsletter: short headed sections, no more than 350 words in total.
Prices and opening hours change. If I have not given you a price, hours or a rule in this
chat, write [TO CONFIRM] instead of guessing. Never invent member names or events.
Ask me one question if the request is unclear.
That runs to roughly 800 characters, well inside the limit. Notice what it does not contain: prices, hours or member names. Those change, so they go in each request or in a project file.
What the output might look like
Prompt: "Reply to someone asking whether they can try a day before committing to a month." Illustrative reply under those instructions:
Yes, you can try a single day first. Book it on our site or tell me the date and I will set it up. A day pass is [TO CONFIRM: price], and if you decide to join within a week we can take it off your first month. Let me know which day suits you.
Check it before sending. The offer to take the day pass off the first month is a promise the model made up, because the instructions did not authorise it. Delete it or confirm it is a real policy. This is the same step as for any AI draft; the instructions reduce errors but do not remove the need to read.
What do custom instructions look like for a sports club secretary?
A sports club secretary can write one set that serves both parents who skim on a phone and volunteers who need to know exactly what to do and when.
The job: weekly emails to parents and volunteers, fixture notices and a monthly update to the committee. The readers are parents with little time and volunteers who want clear tasks.
I am the secretary of a community sports club with about 120 junior and adult members.
I write weekly parent emails, volunteer rotas, fixture notices and committee updates.
Readers: parents (skim on a phone), volunteers (want to know exactly what to do and when),
and committee members (want decisions and numbers).
Voice: warm, brisk, practical. Short sentences. No jargon. British spelling.
Formats:
- Parent email: subject line with the day and event, then bullets: when, where, what to bring.
- Volunteer request: one sentence of why, then the task, the time and who to reply to.
- Committee update: decisions needed first, then figures, then everything else.
Write times as "9:30 am", never in 24-hour form.
Never guess dates, venues, prices or names. If any of those are missing, list what you need
from me before drafting. Do not mention children by name.
The last line is a safeguard worth copying: it keeps junior members' names out of drafts. Put the date, venue and kit list in the request itself each week. A sample request: "Parent email for Saturday 14th, 9:30 am, Riverside pitch 2, bring shin pads and water, under-10s." The instructions supply the format and voice; the request supplies the facts.
How do you write project instructions for catering quotes?
Catering quotes work best as a project whose instructions say to use only the menu and price list files, and to write [NOT IN PRICE LIST] rather than estimate a missing price.
This one is a project instruction rather than a personal set, because quoting is a specific job. Create a project called "Catering quotes" and paste this into Project settings. The project can also hold your menu and price list as files, which the chat can draw on.
This project is for quotes, menus and follow-up emails for a small catering company
serving offices, weddings and private parties.
Use only the menu and price list files in this project for dishes and prices.
If a dish or price is not in the files, write [NOT IN PRICE LIST] and do not estimate.
Quotes: a short intro paragraph, then a table (item, quantity, price per head, total),
then terms in three bullets. Show the per-head price and the total for the headcount.
Allergens: never state that a dish is free of any allergen. Say "check the allergen
sheet" and list the allergens the file shows for that dish.
Follow-ups: under 100 words, one clear next step, friendly but businesslike.
British spelling. No exclamation marks.
Two choices here deserve a comment. First, the allergen rule is deliberately cautious: a wrong "nut-free" statement is a safety risk, so the instructions route the model back to the source sheet. Second, the project reads from files. OpenAI's projects page says you can add PDFs, spreadsheets and documents, and that project instructions override your global custom instructions inside the project. How many files you can upload depends on your plan, so check the limit before you upload a long archive. Business, Enterprise and Edu projects support up to 40 files, per the help page.
A quick sum to check a quote
Ask for a quote for 36 covers at a price per head of $22.50 from your file (illustrative figure). The instructions ask for per-head and total, so you can check it yourself: 36 times $22.50 is $810. If the table shows anything else, the model mis-multiplied or used another figure. Always recompute the total.
Why each line of the coworking set is there
Copying a set without knowing what each line does makes it hard to fix later. Here is the first example annotated, so you can see the thinking and apply it to your own.
- "About 60 members, mostly freelancers and small teams." This sets the scale and the audience. Without it, the model may write as if for a large corporate office with a reception desk and a legal department.
- "They want the answer first." This one sentence changes the structure of most replies, because it tells the model what the reader values over politeness padding.
- "No exclamation marks, no corporate phrases." It names the habit to avoid with one example, "we are delighted to inform you". A concrete banned phrase works better than "be less formal".
- "Replies under 120 words." A number is testable. "Short" is not. Count a few replies and see whether the limit holds.
- "Write [TO CONFIRM] instead of guessing." A marker you can search for is better than a vague "be careful", because you can press Find in the finished draft and see every gap at once.
- "Ask me one question if the request is unclear." Limiting it to one stops the model replying with a list of ten questions, which is its own kind of unhelpful.
Same request, with and without instructions
A before and after makes the value visible. The request is the same: "Reply to a member who says the meeting room was double-booked yesterday." Both answers are illustrative.
Without instructions:
Dear Member, We sincerely apologise for the inconvenience caused by the double booking of our meeting room. We value your business greatly and are committed to providing the best possible experience. Please accept our apologies and rest assured that we will investigate this matter thoroughly. Kind regards, The Team.
With the coworking instructions:
I am sorry about the double booking yesterday. That should not have happened. I am checking how it got through and will message you by the end of tomorrow. As a first step, tell me which slot you needed next week and I will hold it for you.
The second is 52 words, leads with the apology, commits to one concrete step and avoids the stock phrases. It still needs your judgement: "I will message you by the end of tomorrow" is a promise you must be able to keep, and the offer to hold a slot should match your actual process. Edit before sending.
Adapting the pattern to other jobs
The five-block shape fits most small-business roles. A few starting points, with what to change.
- Bookkeeping and admin assistant. Who I am: a small business owner who handles invoices and supplier emails. Format rules: supplier emails under 100 words, amounts always with a currency symbol, dates in the form 14 Feb 2027. Safety rule: never state that a payment has been made or received unless I have said so.
- Retail shop owner writing product descriptions. Voice: warm and specific, two short sentences per product. Format: first line is the product name and material, then one benefit. Safety rule: do not invent materials, sizes, origin or care claims that I have not supplied.
- Event venue sales manager. Readers: someone planning a private event with a budget. Format: lead with capacity and price range, then one question about date. Safety rule: availability is never confirmed in a draft; write [CHECK CALENDAR].
- Clinic or salon front desk. Keep it to booking logistics. Safety rule: do not give health advice or comment on symptoms; refer those questions to a person.
In every case the pattern holds: a short who-I-am, one reader, a voice with one banned habit, formats for the jobs you repeat, and a rule that makes the model ask rather than guess. If you want help turning your own jobs into sets like these, the tutorial on writing AI prompts that sound like your business works well as a companion.
How do you keep a version log for your instructions?
Keep a version log that notes the date, the change and the test prompt behind each edit, because instructions drift when someone fixes one problem and breaks another.
2026-11-17 v1 First set. Five blocks, 820 characters.
2026-11-24 v2 Added "replies under 120 words" after long replies to enquiries.
2026-12-01 v3 Added [TO CONFIRM] rule after an invented day-pass price in a reply.
2026-12-15 v4 Removed three banned phrases that made replies stilted.
The log also tells a new colleague why each rule exists, which is the difference between a rule they respect and one they delete. Review it quarterly, and if a rule has not been relevant for a quarter, remove it.
If you keep that log in a spreadsheet, try The Complete Excel Course with AI Integration on Coding Liquids.
What should you leave out of custom instructions?
Leave out personal data and secrets, and any prices, hours and policies that change, because a longer prompt is not a better prompt. The items below belong elsewhere, or nowhere.
- Personal data and secrets. No passwords, account numbers, customer names or contact details. Custom instructions are applied to every chat and are included in your data export.
- Prices, hours and policies that change. Put them in each request or in a project file you update, not in a personal set you forget.
- Contradictory rules. "Be concise" and "be thorough" in one set produce unpredictable results. Choose one default and override it in the request.
- Legal or medical instructions. "Always give a definite answer on contract terms" is a dangerous rule. Ask for options and have a qualified adviser decide.
- Long lists of banned words. Three banned habits is plenty. A list of fifty makes the writing stilted.
- Anything you would not want a colleague to read. A shared project shows its instructions to collaborators.
OpenAI's page also notes that information from your use of custom instructions may be used to improve model performance, with a link explaining how to opt out, and that if you use third-party plug-ins the model may pass relevant information from your instructions to the plug-in developers. For a Business workspace, OpenAI's Business help says workspace data is excluded from training by default. Even so, the simplest policy is to write instructions you would be comfortable seeing on a whiteboard.
How do you test custom instructions before relying on them?
Test a set with five prompts and compare the answers with a baseline before you rely on it, because a set that reads well can still behave badly.
- The routine job. Ask for your most common task, such as an enquiry reply. Is it the right length and tone?
- The missing fact. Ask for something that needs a price you have not provided. It should say [TO CONFIRM] or ask a question, not guess.
- The off-brand bait. Ask for "something punchy with plenty of exclamation marks". A good set should still follow the voice rule, or push back.
- The awkward case. A complaint, a refund request or an allergy question. Does it stay calm and avoid promises?
- The unrelated task. Ask something else, like a recipe. Do your instructions leak in oddly, such as signing off with your business name?
Create a baseline with a temporary chat
Per OpenAI's help centre (checked 21 November 2026), you choose Personalized or Unpersonalized before a temporary chat's first message; personalisation is on by default, and an unpersonalised temporary chat does not use memory, plugins or custom instructions, so pick Unpersonalized for the baseline. Run the same five prompts there and in a normal chat. Where the answers differ, the instructions are doing something; where they are identical, a rule is being ignored or is too vague to matter. You can also switch off Enable customization briefly for the same comparison.
Score each answer
| Check | Pass looks like |
|---|---|
| Length | Within your stated limit |
| Voice | No banned phrases, right spelling |
| Facts | Gaps marked, nothing invented |
| Structure | Answer first, one next step |
| Safety | No promises, no allergen or legal claims |
Give each answer a pass or fail per row. Two or more fails in a row means the instruction behind it needs rewriting. Change one line at a time, re-run only that prompt, and note the date next to the version.
Fixing common problems
It ignores a rule
Put the rule near the top, state it positively ("write times as 9:30 am") and give one example. Rules buried in the middle of a long paragraph are the ones skipped. Also check whether a project's instructions override it: a conflicting rule in the project wins.
It sounds the same every time
Over-specified voice rules make writing stiff. Replace a paragraph of tone rules with two short sample sentences in your voice and the instruction "match this rhythm".
It refuses to commit to anything
If every answer is full of [TO CONFIRM], your safety rule is too broad or you are not supplying the facts. Narrow it to "prices, dates, names and policies" and give the facts in the request.
Two staff get different results
Their personal sets differ, which is why shared jobs belong in a shared project. Move the job-specific rules into the project, and ask staff to keep personal sets short and general. For comparisons of how other assistants handle this, see the tutorial on what ChatGPT, Claude, Gemini and Copilot remember at work.
How can a small team roll out custom instructions in a week?
A small team can roll instructions out in a week by picking the one job that wastes the most time, writing project instructions for it and testing before sharing the project.
- Day one: list the five jobs your team gives to ChatGPT most. Pick the one that wastes the most time.
- Day two: write the project instructions for that job, using the five blocks above, and add the reference files (price list, policy, past good examples).
- Day three: run the five-prompt test, with a temporary chat as the baseline. Fix two lines.
- Day four: have one colleague use it for real work and note anything they had to correct.
- Day five: share the project with the team, with chat access for most people and edit access for one owner, so rules do not drift.
- Next month: review the corrections list and prune any rule nobody needed.
The brand rules behind your instructions matter as much as the instructions. If you have not written them down, start with the tutorial on building a brand voice guide that AI can follow and paste a short version into your instructions. Menu names and limits change, and OpenAI updated the instruction limit this summer, so check the current help page before you plan around a number.
If you want help fitting a set like this into how your business actually runs, I offer one-to-one AI consultations for small businesses.
Custom instructions: practical questions
How long can ChatGPT custom instructions be?
OpenAI's help says Free and Go users can save up to 1,500 characters, while Plus, Pro, Business, Enterprise and Education users can save up to 5,000. A good set is rarely more than 1,500 characters anyway, because long instructions give the model more rules to trip over.
Do custom instructions or project instructions win?
Project instructions. OpenAI says they apply only within that project and override your global custom instructions. So keep your personal set short and general, and put job-specific rules in the project that does the job.
Can I share my instructions with the team?
Yes, through a shared project: ChatGPT draws on everything in the shared project, including its instructions. I could not find a documented workspace-wide instruction field for Business, so check Workspace settings in your own account before assuming one exists.
Should I put customer details in instructions?
No. Keep names, contact details, passwords, account numbers and anything confidential out of instructions. They are applied to every chat, they are included in your data export, and a shared project shows them to collaborators. Put reusable facts about the business in, and keep personal data out.
Further reads
- How Agencies Keep Brand Voice Consistent in AI-Written Content — How agencies hold one voice across writers.
- ChatGPT Custom Instructions for Consultants: What to Set — Instruction set for a consultancy with data rules.
- How to Write Instagram Captions With AI That Sound Like You — Captions that sound like you, not a template.
- Brand Voice Tools Compared: Jasper, Writer and Custom Instructions — Jasper, Writer and custom instructions compared.
- Does ChatGPT Remember Conversations? Memory Settings for Business — How ChatGPT memory interacts with your instructions.
- How to Set Up a Claude Project for Marketing: Voice and Calendar — The same idea in Claude Projects.
- How to Use ChatGPT Projects to Keep Client Work Separate — Keep each client's work in its own Project.
Browse all AI-for-business tutorials
Sources: OpenAI Help Center articles on ChatGPT custom instructions, projects in ChatGPT, the custom GPT retirement FAQ, ChatGPT Business privacy, and the ChatGPT release notes (27 August and 15 July 2026 entries), read on 17 November 2026.
