How to Set Up ChatGPT for Work: 4 Settings Most People Skip

Why unconfigured ChatGPT sounds generic

ChatGPT writes for the reader it can infer. With no configuration, the only reader it can infer is a generic member of the public, so it produces prose calibrated for a generic member of the public — cautious, explanatory, and padded with context you already have.

This is also why the same prompt gives you different-shaped answers on different days. Nothing is holding the output steady, so small differences in phrasing swing the result. If that inconsistency is your main frustration, the underlying mechanics are worth understanding — we cover them in why ChatGPT gives inconsistent results.

Same model, same request. The difference is that the second account had already been told who is writing, who is reading, and what a finished answer looks like.

There is a quieter cost too. When you have not taken the time to set up ChatGPT for work, you compensate by writing longer prompts — two sentences of role, two of context, one about format — every single time. Thirty seconds of typing, several times a day, forever. The setup below moves that text somewhere it only has to exist once.

Set up ChatGPT for work: four settings

To set up ChatGPT for work properly, you configure four things once: your custom instructions, your memory, your data controls, and your project scope. Everything else is prompting.

The Baseline Four is the approach we teach at PromptPeakAI: four account-level settings that form the baseline every new chat starts from, rather than context you retype each session. They stack — instructions set the standing brief, memory adds what it has picked up, data controls decide what leaves your account, and projects contain the whole thing inside one piece of work.

Work through them in that order. Instructions give you the biggest quality jump for the least effort, and the later settings make more sense once the first one is in place.

Setting 1: your standing instructions

Custom instructions are a standing brief that applies to your conversations automatically, so you stop opening every chat by explaining your job. They are available on every plan, on web, desktop, iOS and Android, according to the official OpenAI Custom Instructions documentation.

On web or desktop, click your profile icon, choose Settings, then Personalization, and make sure Enable customization is toggled on. On iOS and Android the entry is named Customize ChatGPT instead — same settings, different label. If your panel shows separate fields for a nickname, your occupation or preferred traits, fill those too; they feed the same brief.

The most common mistake here is writing a personality profile. What actually changes output is the professional context: your role, who reads your work, and the constraints that never change. Keep it factual and specific.

I am a [job title] in [team or department] at a [industry] company of roughly [size].

My work mostly goes to: [audience — e.g. senior leadership, external clients, my direct reports].

Constraints that always apply:
- We use British English spelling.
- Our regulator is [regulator], so nothing can promise an outcome.
- I write for people who are busy and skim.

Recurring outputs I ask for: status updates, stakeholder emails, meeting summaries, short business cases.

The second half of the brief is where you kill the fluff. Be blunt — negative constraints work better here than polite preferences, because you are overriding default behaviour rather than nudging it.

Default to the shortest answer that fully does the job.

Never open with a compliment, a restatement of my question, or "Great question".
Never close with an offer to help further or a disclaimer unless I ask for risk framing.
Never use the words: leverage, delve, seamless, robust, holistic, journey.

Formatting rules:
- Anything with three or more comparable items goes in a table.
- Anything sequential goes in a numbered list.
- Prose paragraphs stay under 60 words.

If my request is missing something you need, ask one clarifying question before drafting.

Those two halves map cleanly onto the four-part input model we use across the site — Role, Goal, Context and Format, explained in full in the Logic System guide. The instructions field is simply where you park the Role, Context and Format elements permanently, so each prompt only has to carry the Goal.

One practical limit: Free and Go accounts save up to 1,500 characters of instructions, while Plus, Pro, Business, Enterprise and Education accounts save up to 5,000. Free is genuinely enough for most people — 1,500 characters is a tight brief, not a crippled one. If you are weighing the upgrade for other reasons, our breakdown of ChatGPT Free versus Plus covers what else changes. For a deeper tour of the field itself, see our guide to ChatGPT custom instructions.

Setting 2: what memory keeps

Memory is what ChatGPT picks up from your conversations on its own, as opposed to what you deliberately told it in your instructions. Left alone it drifts, and drifted memory is the reason an account starts confidently referencing a project you finished last spring.

This is the part of ChatGPT setup that most guides get wrong, because it changed. The default is no longer a tidy list of saved snippets you delete one by one. It is a memory summary — an automatically maintained synthesis of what ChatGPT has concluded about you, sitting under Settings, then Personalization, then Memory.

You correct it by typing the change into the box at the bottom of the summary, or by highlighting a specific line and fixing it. The three-dot menu also offers a combined delete-and-turn-off option. All of this is documented in the official OpenAI Memory FAQ, which is also where the older behaviour is now labelled as legacy.

If you preferred the old list, it has not disappeared — you can revert to the legacy saved-memories system from the memory settings, and switch back later. Neither version is a vault, though. Fully removing something ChatGPT knows means clearing every source it appears in, including past chats, archived chats and uploaded files.

Before we start: list everything you currently believe about my job, my employer, my active projects and my writing preferences.

Present it as a plain bulleted list.

Flag anything that came from a conversation more than three months old.

Do not add anything you are inferring rather than remembering.

Our fuller walkthrough of how ChatGPT memory works goes deeper on what it stores and why.

Setting 3: your data controls

Turning memory off does not stop your content being used to improve OpenAI’s models. Those are two separate switches, and confusing them is the single most common privacy mistake in professional accounts.

On web, click your profile icon, choose Settings, go to Data controls, and turn off Improve the model for everyone. On mobile, open the sidebar, tap your profile icon, then Data controls. The setting applies to your whole account, so you only do it once. Your chats still appear in your history — they simply stop feeding model training.

Temporary Chat is the icon at the top right of the chat screen. Those conversations do not appear in your history, do not use or create memories, and are not used to train models; OpenAI deletes them after thirty days. It is the right tool for a one-off analysis of something confidential.

Worth knowing if you are comparing plans: on Business and Enterprise workspaces, content is not used for training by default, so the toggle above is a consumer-plan concern. What none of these settings change is the judgement call about what you paste in the first place.

If your organisation has a formal AI policy, read it before you rely on any of the above — our guide to whether ChatGPT is safe to use at work covers the questions worth asking your IT team.

Setting 4: scope work into projects

Projects are containers. Each one holds its own chats, files and instructions, and can be set to project-only memory so that what ChatGPT learns inside it does not leak into the rest of your account. They are available on free and paid plans alike.

This is the setting consultants and anyone juggling clients should reach for first, because it solves a problem your account-level instructions cannot: keeping two similar pieces of work genuinely separate. Use a project when:

• You are working across several sessions on one deliverable
• Two clients or business units need to stay strictly apart
• The same reference files come up again and again
• The work is sensitive enough that you want its context contained

Everything in this project relates only to [client or workstream name].

Do not reference other clients, other projects, or anything outside the files attached here.

House style for this client: [tone, e.g. formal, no contractions, British English].
Their audience: [who reads the deliverables].
Their non-negotiables: [e.g. never quote timelines without a caveat].

If I ask something that needs information not in this project, say so rather than guessing.

The line to hold is this: account instructions describe you, project instructions describe the work. Your job title and writing standards belong at the account level because they travel everywhere. A client’s tone of voice, their banned terminology and their reference documents belong in the project, because you do not want any of it surfacing while you draft something for someone else.

Business, Enterprise and Education workspaces can additionally share projects with teammates, which turns the instructions above into a shared house style rather than a personal one. Our guide to using ChatGPT Projects covers the collaboration side properly.

Which instructions win a conflict

Once you have account instructions, project instructions and possibly a custom GPT in play, they can disagree — and knowing which one wins saves a lot of confused debugging.

The practical consequence: if you have carefully tuned your account instructions and then open a custom GPT, do not expect your tone rules to follow you in. A custom GPT carries its own brief, which is exactly why building one is worth it for genuinely repeatable tasks — covered in our walkthrough on how to build a custom GPT.

What this setup cannot fix

A configured account is a better starting point, not a finished deliverable. It changes the shape and tone of what comes back — it does not change whether the underlying claims are true, and it will state a wrong figure in your house style just as fluently as a right one.

Most managers will also tell you the honest answer to “will my team notice I used AI” is that they notice the generic version, not the configured one. The fix for that is the same fix as everything else here — a brief specific enough that the output could only have come from your desk.

When your setup stops working

Settings that worked in January can quietly stop working by June. Almost always it is one of four things.

One more worth naming: if output quality drops but nothing in your settings changed, the problem is usually the prompt rather than the account. Our method for diagnosing and refining AI output isolates which element actually caused the gap instead of starting over.

The five-minute monthly audit

Configuration decays. Roles change, projects end, and a summary written around your old job keeps shaping answers long after the job is gone. Put five minutes in your calendar once a month.

Review the last five things you helped me write.

Tell me, in plain language:
- What role and audience you assumed each time.
- Which of my standing instructions you actually applied.
- Anything you believe about my work that now looks out of date.

Be specific and do not flatter the output.

Fifteen minutes to configure, five minutes a month to maintain. Against that, most professionals spend longer than fifteen minutes editing a single unconfigured draft into something sendable. If you only do one of the four today, do the instructions — it changes every reply you get for the rest of the week, and it takes about six minutes.

None of the four settings is clever, and none of them is hidden. They are simply the part of the work that happens before the prompt — which is exactly why so few people set up ChatGPT for work at all, and exactly why the ones who do get noticeably better output from the same tool.