How to Set Up ChatGPT Custom Instructions for Professional Work

Why ChatGPT Custom Instructions for Professional Work Matter

Custom Instructions are a block of standing text that ChatGPT reads before every message you send, in every new chat, until you change it. Nothing about the feature is clever. What makes it valuable is that it is the only setting that applies everywhere by default.

That matters because inconsistent AI output is usually an input problem, not a model problem. Two people asking the same question get different answers because they supplied different context. Standing instructions remove that variable permanently — the context stops depending on whether you remembered to type it.

The four layers stack rather than compete, which is why a tone rule sometimes seems to be ignored. If your preset voice is warm and chatty and your instructions say “be blunt”, you have given ChatGPT two jobs. Pick the preset that already points the right way — Professional, Efficient or Candid all suit work — and let your instructions handle the specifics.

Most professionals already have a version of this conversation with new colleagues. You brief them once on what you do, who reads your work and how you like documents laid out, and then you stop repeating it. This is that briefing, written down. If you have not done the wider ChatGPT setup for work yet, do this piece first — it changes every session that follows.

How to set up ChatGPT Custom Instructions

On the web and desktop apps, open Settings and select Personalization. On iOS and Android, open Settings and select Customize ChatGPT. Both routes lead to the same place, and whatever you save applies immediately to new chats.

One honest note before you go looking. OpenAI has reworked this panel several times, and what you see depends on your platform and which rollout you are on. Some accounts show a single instructions box; others show two labelled fields asking what ChatGPT should know about you and how it should respond. The content you write is identical either way, so this guide is organised around the two jobs your text has to do rather than around whichever boxes are on your screen this month.

Click your profile icon, choose Settings, then Personalization on web or desktop. On mobile, the same section is called Customize ChatGPT. If you have never opened it, expect a preset voice selector at the top and your instructions below it.

There is an Enable customization switch. If it is off, your instructions are saved but ignored — which is the single most common reason people think the feature is broken. Turn it on before you write anything.

This is the part worth thinking about, and the next section gives you a structure for it. For now, know how much room you have. The allowance jumped in July 2026 and depends on your plan.

Those figures come from OpenAI’s Custom Instructions documentation, which also confirms the feature works on every plan across web, desktop, iOS and Android. You do not need to pay for this. The free tier gets you roughly a page of standing context, which is more than most people will ever fill sensibly — see our breakdown of what the paid plans actually add if you are weighing an upgrade for other reasons.

Save your changes, open a fresh conversation and send a real work request. Do not assume it took. Check that the tone matches, the formatting follows your rules, and any words you banned are genuinely absent.

The Standing Orders Method

Standing orders are the rules that apply until someone changes them, as distinct from the specific instructions you give for one job. That distinction is exactly the one people miss here. Your instructions box is not where you describe today’s task — it is where you write the three things that are true of every task.

The Standing Orders Method, the approach we teach at PromptPeakAI, says that ChatGPT Custom Instructions for professional work need all three of the following components, and drift back to generic output when any one of them is missing.

Identity is the one people write badly. “I work in marketing” gives ChatGPT nothing to calibrate against. Compare that with a version a new colleague could actually use.

Constraints are the component almost everyone skips, and they do more work than the other two combined. Telling ChatGPT what to stop doing removes the specific tics that make writing read as machine-made. Output standard then locks the shape: how long, how structured, when a table beats a paragraph. Our guide to specifying output format goes deeper on that third component, and the whole method maps cleanly onto the Role, Goal, Context, Format structure we use for individual prompts.

Here is a neutral skeleton to start from. Fill in the square brackets, delete what does not apply, and keep it under a page.

IDENTITY
I am a [job title] at a [size] [sector] organisation. My main outputs are [three things you write most often]. My work is read by [audience]. I have [number] years of experience and do not need fundamentals explained.

CONSTRAINTS
Never use these words: delve, synergy, leverage, robust, game-changer, transformative, seamlessly.
Do not open with a compliment or a restatement of my question.
Do not add disclaimers unless there is a genuine safety or legal risk.

OUTPUT STANDARD
Short questions get two or three sentences of plain prose.
Analysis and planning tasks get headings and short bullets.
Use a table whenever three or more options are being compared.
Lead with the recommendation, then the reasoning.
Spelling: [UK or US] English.

Copy-ready setups for four roles

Each of these is a complete standing brief sitting comfortably inside the 1,500-character free allowance. Paste one, then spend five minutes replacing the generic details with your own — the specifics are what make it work. If you want more of these, our free AI work templates library carries the wider set.

Operations and project managers

IDENTITY
Senior project manager in professional services, running four to six client projects at once. I write status reports, risk logs, stakeholder updates and escalation briefs. My readers are executive sponsors who scan rather than read.

CONSTRAINTS
Never use: synergy, robust, going forward, circle back, bandwidth as a metaphor for capacity.
Do not pad status updates with commentary. Do not soften a red status.

OUTPUT STANDARD
Open every analysis with a one-line "Recommended action".
Use numbered lists for process and sequence.
Use a table for risk logs, with owner and due date columns.
Structure escalations as: issue, impact, options, recommendation.
Keep status updates under 200 words.

HR and people teams

IDENTITY
People Operations Manager at a mid-sized company. I draft internal policies, onboarding material, manager guidance and employee communications. Everything I write is read by staff, so tone carries as much weight as accuracy.

CONSTRAINTS
Never use: delve, strictly prohibited, punitive, failure to comply, synergy.
Do not write anything that reads as a legal threat to an employee.
Do not present a policy draft as legally reviewed — flag where review is needed.

OUTPUT STANDARD
Warm, clear and direct. Plain English over policy language.
Use headings and short bullets for anything over 300 words.
End policy communications with a "Questions and support" section.
Give me the draft first, then a short note on what a lawyer should check.

Marketing and content

IDENTITY
Senior marketing manager at a B2B software company selling to finance and operations leaders. I produce campaign briefs, long-form content, email sequences and monthly performance summaries. My audience is senior, sceptical and short of time.

CONSTRAINTS
Never use: game-changer, unlock, supercharge, delve, testament, revolutionary, seamlessly, cutting-edge.
No exclamation marks. No opening flattery.
Do not invent statistics, customer quotes or case study numbers.

OUTPUT STANDARD
Give two options for any headline, subject line or call to action.
Use a table when comparing messaging angles.
Lead with the recommendation, then justify it.
Keep meta descriptions under 155 characters.

Sales and account management

IDENTITY
Account executive selling technical software to IT and security leaders. I write cold outreach, follow-ups, proposals and internal deal notes. My prospects are time-poor and immediately dismiss anything that reads as a mass email.

CONSTRAINTS
Never use: circle back, touch base, exciting opportunity, game-changer, synergy, just checking in.
No flattery openers. No more than one question per email.
Do not claim results or references I have not given you.

OUTPUT STANDARD
Cold outreach: under 125 words, one specific hook, one low-friction ask.
Follow-ups: under 80 words, and each one must add something new.
Proposals: their problem first, our approach second, us last.
Always give me a subject line under 50 characters.

What never belongs in your instructions

Standing instructions apply globally and sit in your account, so treat them as something a stranger might one day read. That rules out a specific category of content rather than making the feature risky in general.

The plan you are on decides what happens to that text. According to OpenAI’s page on how your data is used, content from individual plans may be used to improve its models unless you opt out through Data Controls, while Business and Enterprise workspaces are not trained on by default. This is worth knowing precisely, because plenty of guides state the opposite.

None of this is a reason to avoid the feature. It is a reason to keep the content at the level of “how I work” rather than “what my company is working on” — which, conveniently, is also what makes instructions age well. For the wider picture on company data, see our guide to using ChatGPT safely at work.

When your instructions stop working

Most failures here are not the model ignoring you. They are one of four specific, fixable causes.

The fifth cause is self-inflicted: too many rules. Thirty instructions that partly contradict each other produce worse output than eight that agree, because you have asked ChatGPT to arbitrate rather than follow. Vague rules make this worse — they read as suggestions.

When something feels off, test rather than guess. Open a new chat and run this.

Without changing anything, tell me:

1. What standing instructions are currently applying to this conversation, summarised in your own words.
2. Which of them are specific enough to act on, and which are too vague to change your behaviour.
3. Any two rules that pull in different directions.

Do not be diplomatic about it.

Instructions, Memory, Projects or GPTs

Custom Instructions are the right tool for things that are true of all your work. The moment a rule is only true sometimes, it belongs somewhere else — and putting it in the wrong place is why setups get bloated.

In practice the split is clean. Your job title and formatting rules go in standing instructions. The fact that you prefer a particular spreadsheet layout is something Memory will pick up on its own. One client’s tone of voice belongs in a Project, because it should not follow you into unrelated work. A repeatable multi-step job with its own reference files has outgrown all three and wants a Custom GPT.

What Custom Instructions cannot do

This is where most guides stop being useful, so it is worth being blunt about the limits. Standing instructions change how ChatGPT writes. They do not change what it knows or whether it is correct.

There is also a maintenance cost nobody mentions. Instructions written for a role you held eighteen months ago quietly degrade every answer, because ChatGPT keeps treating stale context as current. Fifteen minutes once a quarter to re-read and prune is the whole job.

Start with the skeleton, run the diagnostic check after a week of real use, and delete anything that has not visibly changed an answer. Good ChatGPT Custom Instructions for professional work are short, specific and maintained — a tight setup you actually revisit beats an elaborate one you wrote once and forgot. If this is your first piece of standing configuration, the natural next step is building a personal library of the prompts you reuse on top of it.