ChatGPT Custom Instructions Examples for Work: 7 You Can Copy

Stop introducing yourself in every chat

You open a new chat and type the same things again: what you do, who reads your work, that you want the answer first and no pep talk at the end. Custom instructions exist so you write that once. Most ChatGPT custom instructions examples for work that you find online fall into two camps, though. Either they are a single line ("Be concise and professional"), or they are a wall of text that tries to turn the model into a different product.

Neither helps much. Below is a structure that fits in the space you get, seven examples by role that you can copy and edit, and a way to test whether the instructions changed anything.

Where custom instructions live, as of October 2026

In ChatGPT, open Settings, then Personalization, then Custom instructions (or "Customize ChatGPT" from the profile menu). You get short fields for a nickname and occupation and two long text boxes: one about you, one about how ChatGPT should respond. OpenAI has renamed these boxes more than once, but the logic stays the same: facts in the first, rules in the second.

The limit is 1,500 characters per box on the free plan. In July 2026 OpenAI raised it to 5,000 for Plus, Pro, and workspace plans. Custom instructions are available on every plan, and they apply to new chats, not to conversations already open.

The same idea exists elsewhere under other names. In Claude it is "Instructions for Claude" in Settings, available on all plans. In Gemini it is "Instructions for Gemini" under Settings and help, on accounts that have the feature. The examples below work in all three with minor edits.

What belongs there and what doesn't

The most common mistake is putting project-level detail into a global setting. Then the model writes your grocery list in the voice of a B2B proposal.

Layer	What goes in	Example
Custom instructions	Things true in every work chat	Your role, your readers, answer-first, US spelling
Project instructions	Things true for one client or one recurring task	Client background, the format of a weekly report
The prompt	Things true today	This email, this deadline, this document

If a rule is not true in at least nine chats out of ten, it does not belong in custom instructions. Put it in a project instead. We describe that setup in How to create a custom GPT for work, including what replaced the GPT builder.

The structure: facts first, then checkable rules

Box one answers "who is asking." Four or five lines are enough: your role and the kind of company, who reads what you produce, what you produce most often, and any standing facts such as spelling, currency, or units.

Box two answers "what does a good answer look like." Every rule there should be one you can check by looking at the reply. "Be concise" cannot be checked. "Lead with the answer in one or two sentences, then the reasoning" can.

A base version for any professional:

ABOUT ME
I am a [role] at a [size and type of company]. My readers are
[clients / executives / candidates / colleagues outside my field].
Most of what I ask for: emails, short memos, document summaries,
spreadsheet help. US spelling. Amounts in USD.

HOW TO RESPOND
- Lead with the answer in one or two sentences. Reasoning after.
- Default length: under 200 words unless I ask for more.
- If you need a fact I did not give you, ask one question before
  writing. Do not fill the gap with an assumption.
- When you are not sure something is true, say so in the sentence
  where it appears.
- No praise for my question, no summary at the end, no offer of
  further help.
- Drafts I will send to someone: plain sentences, no bullet points,
  no bold, no exclamation marks.

Seven examples by role

Each block replaces the base version above. Edit the specifics in box one; the rules in box two are the part worth keeping.

Marketer

ABOUT ME
B2B marketing manager at a software company, team of four. Readers:
operations and finance buyers who dislike hype. I write campaign
emails, landing page copy, briefs for designers, monthly reports.

HOW TO RESPOND
- Copy: one claim per sentence, each claim tied to something a buyer
  can verify. No superlatives. No invented numbers or customer quotes.
- Give two variants only when I ask for options.
- For reports: the change, the likely cause, what I would do next.
- Mark any statistic you did not get from me as "needs source".

Sales

ABOUT ME
Account executive selling a service with a 2-3 month sales cycle.
Readers: department heads. I write cold emails, follow-ups, call
recaps, and objection replies.

HOW TO RESPOND
- Emails: under 90 words, one ask, no "just checking in".
- Never state anything about the prospect's company that I did not
  give you. Ask instead.
- For objections: restate the concern in their words, then one
  answer. No list of five rebuttals.

Lawyer

ABOUT ME
In-house counsel at a mid-size company. I review commercial
contracts and write to non-lawyers in the business. I paste only
redacted text.

HOW TO RESPOND
- Cite the clause number for every point about a document.
- Separate what the text says from what you infer.
- Never cite a case, statute, or regulation from memory. Say
  "verify" and tell me what to look up.
- Explanations for the business: plain English, under 150 words.

HR

ABOUT ME
HR business partner for a 200-person company. Readers: candidates,
employees, line managers. I write job posts, interview guides,
policy explanations, hard messages.

HOW TO RESPOND
- Neutral, direct, warm without filler. No corporate euphemisms.
- Flag any wording that could read as discriminatory or as a
  promise the company has not made.
- Do not give legal conclusions about employment law. Tell me what
  to confirm with counsel.

Consultant

ABOUT ME
Independent consultant, operations projects for mid-size firms.
Readers: owners and COOs who read on a phone. I write proposals,
call follow-ups, findings memos.

HOW TO RESPOND
- Recommendation first, then at most three supporting points.
- Push back when my reasoning has a gap. Name the gap in one line.
- No frameworks by name unless I ask for one.

Manager

ABOUT ME
I manage a team of eight and report to a VP. I write status updates,
feedback, decision memos, meeting agendas.

HOW TO RESPOND
- Status updates: what changed, what is at risk, what I need. No
  background the reader already has.
- Feedback drafts: describe the behavior and its effect. No
  adjectives about the person.
- For decisions, give me the strongest argument against my option.

Small business owner

ABOUT ME
I own a service business with five employees and do the marketing,
hiring, and client email myself. I have about ten minutes per task.

HOW TO RESPOND
- Give me one recommended option and the reason, not a menu.
- Steps I can finish today come first.
- If something needs a professional (tax, legal), say so in one line
  and continue with what I can do myself.

How the same setup fits into a full working system for a given role is described on our pages for marketers, for sales, for lawyers, and for HR.

Test it: one prompt, before and after

Instructions that you never test are a guess. Pick one ordinary request and run it twice in new chats: once with custom instructions switched off (the toggle sits in the same settings panel) and once with them on.

Draft a reply to a client who asked to move our delivery deadline
out by two weeks. I can agree to one week.

Without instructions, the typical reply is around 200 words. It opens with thanks for reaching out, explains at length why timelines matter, offers the one-week compromise in the third paragraph, and closes with an invitation to "hop on a call."

With the base instructions above, the reply should open with the one-week offer, run under 100 words, and contain no bullets. It may also ask a question first, such as whether the one week is firm. That is the "ask before assuming" rule working.

If you see no difference, the cause is usually one of four things:

1. The chat was opened before you saved the instructions. Start a new one.
2. The rule cannot be checked. Rewrite "be professional" as something visible in the output.
3. Two rules conflict. "Be thorough" and "under 200 words" cannot both win. Delete one.
4. The conversation is long. Late in a long thread, standing instructions lose their grip. Restate the rule or start again.

For answers that are still flat after this, the problem sits in the prompt itself, which we cover in Why ChatGPT gives generic answers.

A six-point check before you save

• Every line in box one is a fact, not a wish.
• Every line in box two can be verified by looking at a reply.
• There is a rule for missing information (ask, or mark the assumption).
• There is a rule for uncertainty.
• Nothing is client-specific or project-specific.
• No names, account numbers, or anything you would not put on a business card.

Where this breaks

Confidential details. Custom instructions are sent with every chat. Keep client names, personal data, and internal figures out of them.

Rules that fit one task and hurt another. "Always answer in bullet points" ruins emails. "Under 200 words" ruins a contract summary. Write defaults with an exit: "unless I ask for more."

Shared projects. A shared ChatGPT project uses project-only memory and does not read your personal custom instructions. Put the rules into the project.

Memory is a separate thing. Memory is what the tool picked up from past chats. Custom instructions are what you wrote on purpose. When they disagree, results get odd. Review saved memories a few times a year.

Voice. Instructions set order, length, and format reliably. They do not make drafts sound like you. That takes samples of your writing and an editing pass, described in How to make AI writing sound human.

Where the practicum picks this up

Persistent settings open session four of our two-week practicum. The session then moves to projects and workspaces and to an assistant for one recurring task, and the homework is done on your own work, so the instructions you end up with have been run against real requests and not sample ones.