Where they live in each assistant
Custom instructions are standing notes the assistant reads before every chat, so you stop retyping “I work in logistics, keep it short, don’t make up numbers” at the start of each one. Every major assistant has them, under different names. Menus move often, so if a path below has changed, search the app’s help for “custom instructions”.
| Assistant | What it is called | Where to find it |
|---|---|---|
| ChatGPT | Custom instructions | Profile menu, then Personalization or Settings. Applies to new chats. OpenAI raised the limit on paid plans from 1,500 to 5,000 characters in July 2026, per its release notes. |
| Claude | Profile preferences, plus project instructions | Settings, then Profile, for every chat. Projects carry their own instructions on paid plans, per Anthropic’s help center. |
| Gemini | Saved info (some accounts say Personal context), plus Gems | Settings, then the saved info page, one line per item. Google’s help page has the current path. |
| Microsoft 365 Copilot | Custom instructions | Settings, then Personalization, where your organization allows it, per Microsoft support. |
Keep each part under 1,500 characters: that fits every ChatGPT account, including free ones, and keeps the text short enough to reuse in the other three. The Custom Instructions builder counts for you and shows where each part goes in each assistant.
The two parts
ChatGPT splits its box into two questions, and they are the right two for every assistant. The first is about you: your job, your organization, the work you do with AI, who reads it and what tools you use. The second is about answers: their shape and length, when to use lists, how to handle facts it is not sure of, and the habits you want gone.
The first part saves you from explaining context. The second part saves you from correcting the same things every day. People usually put effort into the first and leave the second vague (“be concise and professional”), which is backwards. A clear rule such as “Keep answers under 200 words unless I ask for more” changes every answer you get. “Be concise” changes almost nothing.
A complete example
Here is a full set for an operations manager, with each line labeled by the job it does. Point at a label to see its line. Copy the whole thing and replace the details with yours.
I'm an operations manager at a 40-person logistics company.1 I write supplier emails, shift schedules, short reports for the owner and procedures for warehouse staff.2 Most of my readers read on their phones, and for some of them English is a second language.3 We use Microsoft 365 and Excel.4 Answer first, then the reasoning, in plain American English.5 Keep answers under 200 words unless I ask for more. Use bullet points only for steps or lists of three or more items.6 Never invent numbers, names, sources or links. If you are not sure of a fact, say so and tell me how to check it.7 If my request is missing something you need, ask me one question before you answer.8 Do not open with praise for my question and do not end by offering more help.9
- 1RoleWho you are and where you work.
- 2WorkWhat you actually use AI for.
- 3ReadersWho reads what you send.
- 4ToolsSo steps and formulas fit what you have.
- 5ShapeAnswer first, in plain words.
- 6Length and formatA ceiling, and when lists help.
- 7HonestyNo invented facts. Say when unsure.
- 8GapsAsk before guessing.
- 9FillerNo praise up front, no offers at the end.
The whole set is about 750 characters, half the smallest ChatGPT limit. Notice what is not there: no “You are a world-class assistant”, no list of twenty rules, no capitals. Each line is something the model cannot know and that applies to nearly everything this person asks.
What they add to an ordinary request
We tested this the way the Grader tests any prompt. An ordinary one-line request, “Write a reply to this supplier asking why the invoice is higher than the quote”, scores 1 out of 10 on its own. Graded together with the instructions above, the way the assistant reads them, the same line scores 8.
Scores from the Prompt Grader’s free ten checks, run on October 28, 2026, with the instructions graded as a system prompt. The request: “Write a reply to this supplier asking why the invoice is higher than the quote,” followed by the supplier’s email.
The instructions supply six things the one-liner lacked: who the reply is for, background facts, a length, a format rule, what to avoid and what to do when unsure. The two checks they cannot supply are task-specific: what a good reply to this supplier looks like, and an example to match. That is the honest limit of custom instructions. They raise the floor under every request, and the important requests still need a proper brief. Our study of 319 one-line work requests found that 261 scored 1 out of 10, so the floor matters.
Lines worth adding for your job
The example above is general. One or two lines specific to your work usually do more than anything else you could add. Some that pay off:
| If you work in | A line worth adding |
|---|---|
| Management | “My team is eight people in two time zones. Anything written for them should work without a meeting to explain it.” |
| Sales | “Never state prices, discounts or delivery dates unless I give them in the message.” |
| Customer support | “Never promise a refund, credit or date. If a reply needs one, write [CHECK POLICY].” |
| Teaching | “I teach 7th-grade science in the US. Pitch materials at that level, and never ask for or use student names.” |
| Finance | “Show calculations step by step, and mark any figure that did not come from me.” |
| Writing for others | “No long dashes, no ‘I hope this finds you well’, no groups of three adjectives.” |
The last line targets the habits that make text read as machine-written; the guide to AI words to avoid has the full list, and the email guide shows how to add your own voice on top.
What to leave out
- Anything secret. Instructions are stored with your account and apply to every chat. Client names, passwords, salaries and unreleased numbers do not belong there.
- Grand personas. “You are a world-class operations consultant” adds little that your real job title does not already say.
- Rules for one task. “Always format as a table” helps for one report and ruins your emails. Put task rules in the prompt, or in a project.
- Contradictions. “Be brief” and “explain everything thoroughly” in the same box leave the model to pick one each time.
- Shouting. Capitals and “CRITICAL” do not make a rule more likely to be followed, and Anthropic’s guidance warns they can make its recent models apply a rule too widely.
Instructions, projects and memory
Custom instructions are the broadest of three layers. Projects in ChatGPT and Claude, Gems in Gemini and custom GPTs hold instructions for one kind of work, such as a client or a recurring report, and the prompt you type holds the task itself. Each layer should hold only what applies at that level.
Memory is a separate feature. Most assistants can now save details from your chats and use them later. Unlike instructions, which you write, memory is written by the assistant, so it can hold things you did not mean to keep. Look at what it has saved from time to time and delete what is wrong or private, and use a temporary chat for anything sensitive. If you set up a project or Gem for a team, the assistant instructions template writes the instructions and the list of files to attach, and the system prompt guide covers the same job for a customer-facing assistant.
Test them for a week
Instructions are a prompt like any other, so test them. Before you save them, run three requests you make every week in a chat with them and in a temporary chat without, and compare. Then use them for a week and keep a note of every correction you type twice. Those corrections are your next lines. Remove any line you cannot connect to a better answer; a short set the model follows beats a long one it half follows.
A checklist to copy
About you [ ] Your role and organization, in one line [ ] What you actually use AI for [ ] Who reads your work, and how (phone, second language) [ ] The tools you use How to respond [ ] Answer first, then the reasoning [ ] A length ceiling in words [ ] When to use lists, and when not [ ] "Never invent numbers, names, sources or links" [ ] "Say when you are not sure" [ ] "Ask one question if something is missing" [ ] The filler you want gone Before you save [ ] No secrets, client names or passwords [ ] No rules that only fit one task [ ] Each part under 1,500 characters [ ] Tested on three real requests
Key takeaways
- Write two parts: who you are and what your work involves, then how you want answers.
- Rules with numbers and actions (“under 200 words”, “ask one question”) work; adjectives (“concise”) barely do.
- Tell the model never to invent facts and to say when it is unsure. It is the most valuable line.
- Instructions lifted a one-line request from 1 to 8 out of 10, but important tasks still need a full brief.
- Keep secrets and one-task rules out, keep each part under 1,500 characters, and test before you rely on them.
Questions
What should I put in ChatGPT custom instructions?
Two things. About you: your role, your organization, the work you use AI for, who reads it and your tools. How to respond: answer first, a length limit, when to use lists, never invent facts, say when unsure, ask one question if something is missing, and the filler you want gone. The Custom Instructions builder writes both from a few questions.
What is the character limit for ChatGPT custom instructions?
For years it was 1,500 characters per box. According to OpenAI’s release notes, paid plans were raised to 5,000 characters in July 2026. Keeping each part under 1,500 characters means the same text fits any ChatGPT account and also suits Claude, Gemini and Copilot, whose settings are laid out differently.
Do custom instructions apply to chats I already have?
In ChatGPT they apply to new chats, so start a fresh one after you change them. Claude’s profile preferences and Copilot’s custom instructions also apply to conversations going forward. If an answer ignores a new instruction, check that you are in a new chat before rewriting the instruction.
Does Claude have custom instructions?
Yes. Claude calls them profile preferences, under Settings, and they apply to every chat. Projects add their own instructions for one kind of work, and styles change how answers are written. Gemini uses saved info and Gems for the same jobs, and Microsoft 365 Copilot has custom instructions where your organization turns them on.
What is the difference between custom instructions and memory?
You write custom instructions, and they apply to every chat until you change them. Memory is written by the assistant from your conversations, so it can keep details you did not mean to save. Use instructions for anything you want every time, review the memory now and then, and use a temporary chat for sensitive work.
Prompts to use with this
- Write instructions for a Custom GPT, Claude Project or GemYou need a reusable assistant for a repeated task and want clear instructions and a list of files to attach.