ChatGPT Custom Instructions for Consultants: Setup and Data Rules

Set custom instructions that state your role, your usual client type, your preferred structure, your tone and your sourcing rules, in well under ChatGPT's character cap, and add a rule that it asks before guessing. Never paste client names, financials, contracts, credentials or anything under a non-disclosure agreement.

The limit matters less than the content. Free and Go users can save up to 1,500 characters and paid plans up to 5,000, according to OpenAI's help page, and a consultant who fills 5,000 characters with generic advice gets generic output. The generic template in most tutorials is the thing to avoid. A good set is short, specific to how you work, and honest about what the tool must not do.

Follow me on Instagram@sagnikteaches

Where the instructions go and what they apply to

On the web and desktop, OpenAI's help page says to open Settings, choose Personalization, make sure "Enable customization" is on, and enter text in the Custom Instructions field. On iOS and Android the path is Settings, then Customize ChatGPT. The instructions apply immediately to all chats, including existing ones. They are included in your data export, and they are not shared with people who open a shared link to one of your chats.

Connect on LinkedInSagnik Bhattacharya

Two other facts from the same page shape how you should use them. First, OpenAI says information from your use of custom instructions can be used to improve model performance, and links to its opt-out guidance. Second, the page still warns that if you use third-party plug-ins, the model may give their developers relevant information from your instructions. Neither is a reason to avoid the feature, but both are reasons to treat the field as a style sheet rather than a notebook. Anything you would not email to a stranger does not belong in it.

Subscribe on YouTube@sagnikteaches

Global instructions versus a project

Projects in ChatGPT keep related chats, files and instructions together. OpenAI states that project instructions apply only within that project and override your global custom instructions. That gives a clean split for a consultant:

• Global custom instructions: how you work. Role, tone, structure, sourcing rules, what to do when unsure. Same for every client.
• One project per engagement: what you are working on. Sector, goals, terms of reference, file of non-confidential background. Different for each client, so nothing carries over by accident.

The practical benefit of the split is that when an engagement ends you can delete that project, after saving your deliverables elsewhere, and the global set stays untouched.

A global set you can adapt

This is a worked set of about 1,350 characters, so it fits even the Free and Go cap. It is written as plain instructions; the model does not need headings or clever formatting. Replace the bracketed parts with your own.

I am an independent consultant advising [size and type of clients, e.g. owner-managed businesses with 10 to 100 staff] on [your specialism]. My readers are busy owners and managers, not specialists.

Writing: plain English, short sentences, -ise spellings. No jargon without a one-clause definition. No filler openings or closing summaries. Do not use the words "leverage", "synergy" or "robust".

Structure: start with the answer or recommendation in two sentences. Then give reasoning, then risks, then next steps with an owner and a date placeholder. Use a table only when comparing three or more options on the same criteria.

Sourcing: do not invent statistics, studies, quotes or client examples. If you cite a fact, say where it comes from and how current it is. If you cannot verify something, say "unverified" and tell me what to check. Never present an estimate as a fact.

Behaviour: if my request is ambiguous or missing key information (numbers, scope, audience), ask up to three questions before drafting. Challenge weak assumptions in my brief. Give a view, not a list of options, unless I ask for options.

Confidentiality: I will use placeholders such as [Client A], [Sector], [Revenue figure]. If I paste anything that looks like a real client name, personal data, contract wording or credentials, stop and warn me instead of using it.

Each paragraph has a job. The first stops the model writing for the wrong reader. The second and third control output style. The fourth turns off the "yes, and" behaviour that makes AI useless for advice. The last creates a tripwire for the exact mistake consultants fear. It will not catch everything, so it complements the never-paste list below rather than replacing it.

What each section fixes, with an illustrative before and after

Say you ask: "Draft a recommendation on whether a 40-person distribution business should outsource its returns handling." With no instructions, the typical answer is a general essay: "Outsourcing can offer cost savings, but there are also risks to consider..." followed by two bullet lists.

With the set above, the first thing you get back is usually a question or two: "What is the monthly volume of returns, and what does handling cost today per item? Is the business open to a variable-cost contract?" Once you answer, the draft starts with a two-sentence recommendation, followed by reasoning, three risks and next steps. The difference is not intelligence but constraints, and the questions the model asks first are the ones you would ask a client in the first meeting.

Role, tone, structure and sourcing: what to put in each

If you write your own rather than adapting the sample, work through four decisions.

Role

State who you are advising and what kind of problems you solve. "Independent consultant" is too thin. "Independent consultant on operations and process for owner-managed trade and distribution firms" lets the model choose sensible defaults, such as not recommending enterprise software to a twelve-person business. Avoid claiming credentials you do not hold; the model will take them literally and write with misplaced authority.

Tone

Give the tone as behaviours, not adjectives. "Direct, calm, no cheerleading" works better than "professional but friendly". List two or three words you never use. If you have a published article or a proposal you are proud of, you can describe its qualities in a line ("short paragraphs, concrete numbers, one clear ask"), though avoid pasting a client-facing document that carries client detail.

Structure

Consultants live in repeatable formats: executive summary, options, risks, next steps. Fix the order you want and the length of each. If you prefer deliverables in a particular shape (a one-page memo, a decision table), say so, and say when not to use it.

Sourcing

This is the section that protects your reputation. A confident wrong statistic in a client report is a professional failure. Tell the model not to invent figures, to label estimates as estimates and to flag what you must verify. Then do the verifying yourself, because an instruction reduces the problem without removing it. For checking habits, using ChatGPT as a business adviser without being misled goes deeper.

Projects for shared client context

A project is where the specifics live. A sensible structure for an engagement:

1. Name it neutrally. "Client A - process review" is fine; the real client name is not needed in the sidebar of a tool you may share your screen from.
2. Write project instructions of 5 to 10 lines: the objective in one sentence, the audience for deliverables, the sector and size, what is in and out of scope, and the placeholders you will use.
3. Add only non-confidential files. Published sector reports, your own methodology notes, generic templates, an anonymised data extract. Not the client's contract, not their management accounts.
4. Keep a "decisions so far" note that you update by hand at the end of each session, so the next chat starts from the same page.

OpenAI's page says shared projects can draw on chats, uploaded files and custom instructions, and that sharing options depend on your plan and workspace settings. For a solo consultant a project is a personal workspace. If you work with an associate, sharing a project is convenient but widens who sees the context; share only with people who are already bound by the same confidentiality terms, and check the plan and workspace settings first.

File limits depend on the plan. OpenAI's page lists 5 files per project on Free, 25 on Go and Plus, and 40 on Pro, Business, Enterprise and Edu, with no more than 10 uploaded at once. That is plenty for background material, and a sign that the project is meant as a working folder, not an archive of the engagement.

A filled-in project instruction (illustrative)

Project: Client A - returns handling review.
Objective: decide whether to outsource returns handling and, if so, to what kind of provider.
Client: [Sector] business, about 40 staff, owner-managed. Audience for deliverables: the owner and the operations manager.
In scope: options, costs, risks, a recommendation, a 90-day plan. Out of scope: legal advice, tax, IT selection.
Use placeholders: [Client A], [Monthly returns volume], [Cost per return]. I will tell you the real figures in my head, not in this chat.
Style: follow my global instructions. Recommendations first.

The line "I will tell you the real figures in my head, not in this chat" is slightly tongue in cheek, but it sets the habit: you work with ranges and placeholders, then drop the real numbers into the final document yourself.

The never-paste list for client data

Write this down once and pin it where you work. It is stricter than most firms' policies, because the cost of a leak is borne by the client and your reputation, not by the tool.

Add one more rule that is not about content: check the terms of your client agreement. Some clients prohibit uploading their material to third-party AI tools at all, and some require written approval. If an agreement is silent, ask rather than assume. The wider question is covered in whether customer data is safe in ChatGPT.

Account settings that go with the rules

OpenAI documents a data-controls setting for whether your content is used to improve its models, and says business plans such as ChatGPT Business and Enterprise do not train on business content by default. If you are on an individual plan, look at your data controls and decide deliberately. Labels and defaults change, so check the current screen. Also use Temporary Chat for anything sensitive-adjacent, and delete chats you no longer need. Keep in mind that a business workspace has its own admin and retention rules, which your client's own policy may or may not accept.

Five realistic uses, each within the rules

1. Interview guide. You give a placeholder-only description ("[Sector] firm, 40 staff, returns backlog") and ask for ten questions to put to the operations manager. Output: a first draft you tailor. Fix: delete questions that assume a system the client does not have.
2. Option comparison. You supply three options with rounded costs. ChatGPT produces a table of trade-offs. Fix: check every number it adds; reject those you did not give it.
3. Executive summary tightening. You paste your own summary paragraph, with names replaced, and ask for a version under 120 words. Output is usually shorter and flatter. Fix: restore the one insight that made it yours.
4. Workshop plan. You describe a half-day session with eight attendees and ask for a timed agenda. Fix: cut anything that needs props you will not bring.
5. Red-team pass. You describe your recommendation in generic terms and ask "what would a sceptical finance director challenge?" The questions are often the useful part; answers you write yourself.

Notice that none of these needs real client identifiers. If a task does, that task is the one to keep out of ChatGPT, or to run on a plan and under terms your client has agreed.

What to do on a plan with the lower cap

If you are on Free or Go, the 1,500-character limit forces discipline, and the sample set above is written to fit it. If you are trimming your own, cut in this order: first the tone adjectives (keep only the banned words), then the structure detail (keep "answer first, then risks, then next steps"), and last the sourcing and confidentiality rules, which do the most work per character. A compressed version might read: "Independent operations consultant for owner-managed firms. Answer first in two sentences, then reasoning, risks, next steps. Plain English with -ise spellings. Never invent statistics; label unverified claims. Ask up to three questions if my brief is vague. Warn me if I paste names, contracts or credentials." That is under 400 characters and still does the main jobs, which shows how little the cap actually constrains a well-written set.

On a paid plan with the larger cap, resist the urge to use it. Long instruction sets get partly ignored, contradict themselves and become hard to maintain. If you find yourself above 2,500 characters, the extra almost certainly belongs in a project.

Sample output with a fix: the vague brief

Illustrative only. You type: "Write a recommendation for a client on whether to buy new software." With the instructions above, a good response looks like this:

Before I draft, three questions:
1. What problem is the software meant to solve, and how is it handled today?
2. Roughly what budget and timeline does [Client A] have?
3. Who signs off, and what would make them say no?
Placeholders are fine for the figures.

That is the behaviour you want. The failure mode is when it answers anyway with a confident recommendation based on nothing. If that happens, your "ask before drafting" rule is too weak: change it to "Do not draft until I have answered your questions" and run the test again.

Sample output with a fix: the unsourced figure

You ask: "What share of small distributors outsource returns?" A response that follows your sourcing rule might say: "I do not have a figure I can verify. Industry bodies and logistics providers publish surveys on this, but the definitions differ, so treat any single number with caution. Check the latest report from a trade association in your client's sector and note the sample size." What you fix: nothing in the output, but you now know to go and look. The failure mode is a plausible number with a vague attribution ("research suggests 30%"). If you see one, reply "source or remove" and treat the answer with suspicion. Never carry an unsourced figure into a client document.

Mistakes consultants make with custom instructions

• Pasting a client's brand voice guide into the global instructions. It leaks one client's tone into every other piece of work. Put it in that client's project.
• Using the instructions to store facts. "My day rate is..." or "Client A's CFO is..." has no business there. Facts about people and prices go in your own records.
• Writing instructions as a single long paragraph. Short labelled paragraphs are easier to edit and debug, and the model follows them just as well.
• Forgetting the rules exist. An instruction you wrote six months ago may quietly shape every answer. Re-read the set whenever the output feels off, and on a schedule.
• Assuming the tool enforces your data rules. The "warn me if I paste a client name" line is a nudge, not a filter. The never-paste list works only if you follow it.

Working with an associate or a subcontractor

If someone else helps with an engagement, give them the written rules, not a copy of your account. A short briefing note that covers the global instruction set, the never-paste list, the placeholder convention and the approved plan is enough. They should have their own account settings and their own confidentiality terms with you. Where you do share a project, remember from OpenAI's help page that owners and members with edit access can add and remove files and change instructions, so agree who is allowed to change what, and remove access at the end of the work. Two more details from the same page help here. A shared project uses project-only memory and cannot see a member's own custom instructions or memories from outside it, so your global set and your associate's do not mix inside the engagement. And the number of collaborators depends on the owner's plan: OpenAI lists 5 on Free, 10 on Go and Plus, and 100 on Pro.

Telling clients what you do

Clients increasingly ask whether their material is going into AI tools. A one-paragraph statement in your engagement terms answers it before they ask: which tools you use, that you do not upload their confidential material to them without written permission, that you work with anonymised or placeholder data otherwise, and that all output is reviewed by you before it reaches them. That statement is only worth making if your habits match it, which is the point of the never-paste list. Whether to tell customers you use AI covers the wording choices.

Closing an engagement cleanly

When the work ends, delete the project (OpenAI says this permanently removes its chats, its instructions and files stored only in the project), delete any other chats that contained working material, and remove any files you uploaded elsewhere. Keep your deliverables in your own document store, not in ChatGPT. Then note what worked in the project instructions: if a rule or a question proved useful, promote it to your global set so the next engagement starts better. Over a year this is how the global instructions become genuinely yours rather than a template.

How to check that your instructions are working

After saving the instructions, run a three-prompt test in a fresh chat: an ambiguous request ("write me a proposal outline"), a request for a statistic ("how big is the market for this?"), and a deliberately bad paste (a fake client name and fake email). The right behaviour: it asks questions, it declines to supply an unsourced figure or labels it unverified, and it warns you about the paste. If any of those fails, tighten the relevant paragraph rather than adding more text. Re-run the test whenever you change the instructions or the model changes, because behaviour shifts across versions.

Review the set every quarter. Delete anything you have not noticed making a difference, and move client-specific lines into projects. If you want a wider look at whether your practice is ready for AI beyond the chat window, the AI readiness checklist for professional firms is a good next step, and keeping prompts and data portable explains why you should store a copy of your instructions in your own notes, outside any one vendor.

If you would rather build the set and the data rules with someone, that is what my AI implementation consultation covers.

Questions consultants ask about custom instructions

How long can ChatGPT custom instructions be?

OpenAI's help page says Free and Go users can save up to 1,500 characters and Plus, Pro, Business, Enterprise and Education users up to 5,000. Because limits change, check the help page. Aim to use well under the cap so the instructions stay readable and easy to maintain.

Do project instructions replace my custom instructions?

Inside a project, OpenAI says project instructions apply only within that project and override your global custom instructions. So keep the global set short and general, and put client-specific rules in the project where they belong.

Are custom instructions a safe place for client details?

No. Treat them as a style and role sheet, not a store for client information. OpenAI's help page says information from your use of custom instructions is used to improve model performance, links to how to opt out, and warns that third-party plug-ins may receive relevant information from them.

Should I use the same instructions for every client?

Use one global set for how you work, and one project per client or engagement for context and rules. A single set that tries to cover every client either bloats past the limit or leaks one client's assumptions into another's work.

Further reads

• Turning Hours Saved by AI Into Advisory Work Clients Pay For — Turn time saved into paid advisory work.
• Can Consultants Use ChatGPT With Client Data? NDAs and Settings — NDAs, client data and the settings that go with them.
• What an AI Implementation Looks Like in a Small Accounting Firm — A professional-services AI rollout, step by step.
• How to Pilot AI in Shadow Mode Before Customers See It — Trial AI on your own work before client work.
• ChatGPT Prompts for Strategy Consultants: Issue Trees and Hypotheses — Issue-tree and hypothesis prompts once the set-up is done.
• How to Write AI Prompts That Sound Like Your Business — Make prompts sound like your own firm.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: OpenAI Help Center pages on ChatGPT custom instructions and Projects in ChatGPT, read in November 2026.