The Best ChatGPT Prompts and Custom Instructions to Try

You have read the "be specific, give context" advice. You have applied it. And you have probably typed some version of "I'm a marketer, I need this for a technical audience, write it direct, skip the hedging" into a dozen separate conversations this month alone, retyping it each time because ChatGPT forgets you the moment you close the tab. That is the tax. Every conversation starts from zero, and every conversation punishes you with bland, both-sides, "as an AI language model" output when you do not pay it.

Here is what nobody told you. That is not a prompting problem. It is a configuration problem, and you have been solving it at the wrong layer. This article gives you the actual text to paste into ChatGPT's Custom Instructions boxes, matched to what you actually do, plus a line-by-line explanation of why each one changes what you get back. It also gives you a working library of the best prompts for ChatGPT to keep for the tasks that Custom Instructions cannot cover. A five-minute setup you do once, and a set of per-message moves you reuse forever. The focus throughout is on what actually changes the output, and we present each fix with the reasoning behind it so you can adapt it to your own work.

Read: How to Become an AI Specialist

Why ChatGPT Keeps Giving You Generic Answers (It Is Not Your Prompts)

When ChatGPT does not know who you are, it hedges. Not because it is lazy or broken, but because it is doing exactly what it was built to do, which is to produce the response least likely to be wrong for the average person who might have typed your question. People often treat it like a search engine and expect one clean answer, but ChatGPT responses are shaped by whatever context you supply, so with no context you get a safe average instead of the key points you actually wanted.

Think about what that means mechanically. The model runs on machine learning trained across a vast range of data, and it has no idea whether you are a McKinsey consultant aligning a recommendation to your business goals or a college sophomore writing a first paper. Like most AI tools, it can only work with what you feed it. So it splits the difference. It gives you the balanced framing, the "it depends on your specific situation" caveat, the safe middle of the range. That output is built to be acceptable to everyone, which is another way of saying it is useful to no one. The genericness is not a bug in your prompt. It is the predictable result of a model guessing at a reader it cannot see, because it has no fixed target audience until you give it one.

So you compensate. You front-load every message with context, then the conversation ends, ChatGPT discards all of it, and the next time you open a fresh chat you are back to explaining yourself from scratch. This is the re-specification tax, and it compounds silently. The fix for a large share of your bad outputs is not a better prompt written a thousand times. It is a one-time configuration set once, in a place most daily users have never opened. The per-message tips you have already absorbed, which are to be specific, provide context, and assign a role, are not wrong. They are just aimed at the wrong layer of the tool. They tell you to re-supply your context on every message instead of making it permanent.

Read: AI Upskilling: Top Firms, Programs, & Tools for Training Your Workforce

Custom Instructions vs. a One-Off Prompt: What Actually Persists

Custom Instructions are two text boxes where you write, once, the context and rules you want applied to every new conversation automatically. You set them and forget them. ChatGPT reads them at the start of every fresh chat without you ever typing them again.

To find them, click your profile icon, open Settings, then select Personalization and Customize ChatGPT. You will see two fields. The first, "What would you like ChatGPT to know about you?", is for persistent context, meaning who you are, what you do, and who you work for. The second, "How would you like ChatGPT to respond?", is for persistent rules, meaning tone, format, and behaviors you always want. Each field holds roughly 1,500 characters, so you have space but not unlimited room. Both apply automatically to every new conversation you start, and Custom Instructions are now available across all plans rather than paid tiers only. (Last verified against OpenAI documentation and the live interface in 2026. Menu paths and labels change, so confirm in your own Settings.)

One recent change matters here. Above the two boxes, the current interface adds a "Base style and tone" personality dropdown with presets such as Default, Professional, Friendly, Candid, and Efficient. This preset sets the overall tone, and your written instructions layer specific rules on top of it. If your output ever feels off in a way your instructions do not explain, check which preset is selected, because the lower tone layer can quietly pull against the rules you typed.

Now the part that trips people up. Custom Instructions are not the same as ChatGPT's Memory feature, and neither is the same as a one-off prompt. Here is the full distinction.

The strategic point buried in that table is that Custom Instructions are more reliable than Memory for persistent context, precisely because you wrote them. They apply the same way every time, and they contain only what you decided they should. Memory works differently and now runs in two layers. There are "saved memories," an editable list of facts you can view and change, and "reference chat history," a broader layer where ChatGPT draws on patterns from your past conversations rather than a fixed list. That second layer accumulates on its own, which means it can retrieve the wrong context at the wrong moment, or hold onto something that was true three projects ago. If you want ChatGPT to reliably know your role and honor your format rules, that belongs in Custom Instructions, not left to Memory's discretion. You can review and prune both layers under Settings, Personalization, Memory whenever the context feels off. (Last verified against OpenAI's Memory documentation in 2026.)

The single highest-leverage feature in ChatGPT is the one most daily users have never clicked. The person complaining loudest about generic output can usually identify the fix in under a minute once they open the Custom Instructions panel, because they have never once opened it.

Read: How to Use ChatGPT at Work: Prompts That Save Hours Every Week

The Custom Instruction Sets, By Role (Copy, Paste, Adapt)

Below are five complete, paste-ready instruction sets. Find the one closest to your role, drop the text into your two boxes, and edit the specifics such as company, stack, and target job. Under each, an italicized note explains the one or two lines doing the heaviest lifting, so you can adapt, not just copy. The goal is to create instructions that match your specific needs and produce accurate, on-target output, not to collect someone else's.

The Marketer / Content Lead Set

Box 1, What would you like ChatGPT to know about you?

Box 2, How would you like ChatGPT to respond?

Why the audience line carries the set: Naming "technical buyers, engineering managers and VPs of product" once means ChatGPT calibrates register and vocabulary in every conversation without you re-specifying it. Left blank, it defaults to a generic professional voice pitched at nobody. That one line is the difference between copy written for your buyer and copy written for "a business audience." This is the highest-leverage line in the article, because it does the most to fix content creation quality before you type a single task.

The Founder / Operator Set

Box 1, What would you like ChatGPT to know about you?

Box 2, How would you like ChatGPT to respond?

The highest-leverage line for founders is "challenge it, don't flatter me." ChatGPT is trained to be agreeable, and left alone it will validate a mediocre plan because agreement is the statistically safe response. Instructing it to attack the idea first, and to recommend rather than enumerate, overrides that habit and turns it into a thought partner instead of a yes-machine. This mirrors what the sharpest ChatGPT users report, which is that the anti-flattery line changes output quality more than any tone instruction.

The Consultant / Advisor Set

Box 1, What would you like ChatGPT to know about you?

Box 2, How would you like ChatGPT to respond?

The highest-leverage line for consultants tells the model to surface assumptions and mark confidence. By default, ChatGPT states everything with the same flat certainty, so a wild guess and a well-grounded fact arrive in the same confident tone, which quietly corrupts any analysis you build on top of it. Forcing it to separate the two lets you actually trust the output enough to put it in front of a client, because you can see which concepts rest on solid ground and which do not.

The Engineer / Technical Set

Box 1, What would you like ChatGPT to know about you?

Box 2, How would you like ChatGPT to respond?

The highest-leverage line for engineers is "assume I'm technical, skip the caveats." When ChatGPT does not know your level, it defaults to the least-technical plausible reader, which is why it buries working code under paragraphs of "before we begin, it's important to understand" and safety hand-holding, the kind of padding you would strip out of real technical documentation. This line collapses all of that. You stop scrolling past explanations you did not need to reach the two lines of code you asked for.

The Job Seeker Set

Box 1, What would you like ChatGPT to know about you?

Box 2, How would you like ChatGPT to respond?

The highest-leverage line for job seekers is the voice anchor in Box 1, which stops every output from sounding like an AI-generated LinkedIn post. Recruiters screen out generic ChatGPT boilerplate on sight, and the default output is exactly that boilerplate. Describing how you actually write, meaning short sentences, concrete numbers, and no exclamation points, gives the model something specific to imitate instead of reaching for the buzzwords that get applications filtered.

Read: How to Use ChatGPT to Write Your Resume (Prompts + Examples)

The Lines That Do Nothing (What to Delete From Your Instructions)

Half of what people paste into their Custom Instructions is theater. It feels powerful, it reads like a real prompt engineering technique, and it changes nothing about the output. Here is what to strip out, and why each one is inert.

"Act as a world-class expert" or "You are the best [X] in the world"

A superlative gives the model no information it can use. Accuracy does not improve because you called it the best. What actually filters the model's knowledge is naming a specific, real role, such as "content marketing lead" or "software engineer familiar with LLM APIs," which narrows what part of its training it draws from. Describe the real role, not a superlative, because "world-class" narrows nothing.

Elaborate multi-paragraph personas and backstories

The three-paragraph character sketch, such as "You are Alex, a seasoned strategist with 20 years at Fortune 500 firms who speaks with quiet confidence," consumes your instruction budget and buys you a faint tone flavor at best. It does not change substance or correctness.

Threats and incentives

Lines such as "I'll tip you $200," "this is critical to my career," or "my job depends on this" are folk techniques with no reliable, durable effect on current models. Recent empirical testing found these tricks do not hold up, and widely used instruction sets have started removing them for that reason.

Vague quality demands

Lines such as "be amazing," "give me the best possible answer," or "think carefully" do not specify anything the model can operationalize. Compare them to "give me a recommendation, not a list of options," which is a concrete behavior the model can actually execute. "Be amazing" is a wish. The other is an instruction.

The general test, so you can audit any set you find on Reddit or in a listicle, is simple. Keep a line only if you can name the specific behavior or piece of information it changes. If you cannot, if it is just vibes, delete it. The lines people copy most enthusiastically, the dramatic personas and the "world-class expert" openers, are consistently the ones that do the least. That is not a theory about how the model should behave, and it is not something research on the model can suggest in the abstract. It is what shows up when you watch output across dozens of setups with and without those lines.

What Heavy Users Actually Keep in Their Instructions

Theory is one thing. It helps to see what people who run ChatGPT all day actually settle on, because their instructions get stress-tested against real work rather than written for a blog post. A few patterns show up again and again in the communities where power users trade setups.

The most common keeper is some version of the anti-flattery line. Users are tired of the reflexive "great question," and the padding that opens every response, and the instruction they reach for reframes the model from an agreeable assistant into a critical collaborator, telling it to earn any praise rather than hand it out by default. This matches what the role sets above already do, and it is the single most widely shared instruction for a reason.

The second pattern is banning the em dash. It has become the visible tell of unedited AI text, and a lot of people strip it out specifically so their writing does not read as machine-generated. If you write in public, adding "do not use em dashes" is a small line with an outsized payoff.

The third insight is subtler and worth internalizing. Long, wordy instructions tend to produce long, wordy answers, because the model mirrors the structure it is handed. If you pack your boxes with twelve elaborate rules, do not be surprised when responses come back bloated. The users who get consistently tight output keep their instructions short and concrete, which is the same lesson as the "lines that do nothing" section approached from the other side.

One last tip that experienced users swear by. When you are unsure whether an instruction will land, paste your draft into ChatGPT and ask it how it would interpret each line, and where two lines might conflict. It is surprisingly good at spotting the ambiguity and the contradictions you cannot see in your own wording, and it takes thirty seconds.

The Best Prompts for ChatGPT (Because Custom Instructions Can't Do Everything)

Custom Instructions handle who you are and how you always want responses. They do not handle what you need right now. The division is clean. Durable identity, tone, and format go in the boxes, set once. Task-specific context, inputs, and one-off constraints go in the prompt, every message. The most common mistake operators make is cramming task-specific detail into Custom Instructions, which bloats every unrelated conversation with context that only mattered for one project. Keep the boxes to what is always true. Handle the specifics per message.

The best prompts for ChatGPT are not clever incantations. They are structures that reliably shape the output. Understanding why these prompts work, rather than memorizing magic words, is what lets you build your own solutions for new tasks as they come up, the same way you would with any of the AI tools in your stack. Here is a working library, organized by what you are trying to do. Keep the ones that match your work and reuse them.

Show, don't tell: the few-shot example prompt

Show the model what you want instead of describing it.

Before: "Write three punchy subject lines for a product launch email."

After: "Write three subject lines in this style. Examples of the tone I want: 'The update you've been asking for.' / 'We fixed the thing everyone hated.' / 'Faster. Finally.'"

Two or three examples constrain the output far more reliably than any adjective. "Punchy" is a word the model has to interpret. Your three examples are a pattern it can imitate directly. It copies the structure and register of what you present instead of guessing at what "punchy" means to you. The same move works whenever you want new content in a fixed style, and it applies well beyond text, since you can show a reference when you ask ChatGPT to generate an image in a particular look rather than describing the look in words.

Question first: the clarify-before-you-write prompt

Stop the model before it commits to a wrong first draft.

Before: "Write a cold outreach email to a potential partner."

After: "Before you write anything, ask me the three questions you most need answered to make this outreach email land. Then wait for my answers."

Front-loading clarification beats correcting a bad draft. A blind first attempt forces the model to invent the details you did not supply, and you spend the next three messages walking it back. Asking for questions at the beginning costs you fifteen seconds and saves the whole re-drafting spiral. This approach helps most on open-ended tasks where a wrong first draft is expensive to unwind.

Prosecute the idea: the argue-against-yourself prompt

A per-message version of the founder's anti-flattery line, for when you have not set it permanently.

Before: "Is launching this feature next month a good idea?"

After: "Before you answer, argue the strongest case against launching this feature next month. Then give me your actual recommendation."

Forcing the model to build the counterargument first defeats its default agreeableness. Ask it whether your idea is good, and it leans toward yes. Make it analyze the weak points and prosecute the idea before it defends it, and you turn a rubber-stamp into a real pressure-testing process that surfaces the objection you actually needed to hear.

Teach it back: the step-by-step explainer prompt

For when you are learning a complex concept, not producing a deliverable.

Before: "Explain how vector embeddings work."

After: "Explain how vector embeddings work as a step-by-step guide, starting from what problem they solve, using one concrete example, and stopping to check my understanding before each new idea."

Naming the format you want, a step-by-step guide with a worked example, pulls the answer out of lecture mode and into something you can actually follow. This is one of the most useful prompts for anyone using ChatGPT to learn rather than to produce, because it turns a wall of text into a sequence you can absorb. It also makes dense concepts genuinely interesting to work through, and it beats most static resources because you can stop and ask a follow-up the moment something does not click.

Set the frame: the role-plus-constraints prompt

The simplest reliable structure, drawn from what heavy users report works. Define role, then constraints, then output format.

"Act as a senior financial analyst. Be concise, no filler or praise. Use short paragraphs, not bullet lists. If a number is an estimate, label it. If you are unsure, say so. End with one thing I should double-check."

This works because it does three jobs at once. The role narrows the knowledge the model draws on, the constraints kill the default hedging and flattery, and the output format tells it how to structure the answer. It is the per-message twin of a good Custom Instruction set.

One note on model choice

Reasoning models, the current GPT-5-class models with adjustable reasoning effort, behave differently from standard models. They already reason through problems internally before they answer, so the old "think step by step" instruction is redundant and can actually degrade the result by making the model narrate reasoning it already did. OpenAI's own guidance now says to avoid chain-of-thought prompts on these models and to write plain, direct requests instead. Save the step-by-step prompting for standard models, define the goal and constraints rather than the procedure, and let the reasoning models reason. It is worth a quick experiment across both model types on your own tasks, since the guidance here will keep shifting as future models change how they handle reasoning. (Last verified against OpenAI's reasoning-model guidance in 2026.)

How to Test Whether Your Custom Instructions Are Actually Working

Do not trust that your instructions took. Verify it in two minutes. Run one ordinary prompt you would normally type, such as "summarize this article" or "draft a quick reply to this email," in a chat before you set your instructions. Then set them, open a fresh chat, and run the identical prompt on the same subject. Compare the two. You should see the preamble disappear, the tone shift to what you specified, and the unrequested hedging drop out. If those three things changed, it is working. This quick before-and-after scenario is the fastest way to test the effect for yourself rather than assuming it landed, and it gives you immediate feedback on which lines are pulling their weight.

If it seems like nothing changed, the culprit is almost always the same one. Custom Instructions apply to new conversations, not the one you already have open. If you set your instructions and then keep typing in the same chat you were in, you will see no difference and conclude the whole thing failed. Start a new chat and test there. This one gotcha accounts for the large majority of "my instructions don't work" complaints.

If it is still not landing, check these in order.

• New chat? Confirm you are testing in a fresh conversation, not a continued one.
• Conflicting rules? If Box 2 says "be terse" and also "explain your reasoning thoroughly," you have handed the model two contradictory instructions. Reconcile them, and decide whether you want brevity or depth as the default.
• Personality preset fighting you? Check the "Base style and tone" dropdown. A preset like Friendly can soften the direct tone you asked for in the box.
• Too long? Instructions that run to a wall of text dilute themselves, and the model cannot prioritize twelve competing rules. Keep each box tight and ordered by what matters most.
• Which model? Behavior varies across versions, and an update can change how instructions get interpreted. Confirm which model you are using before assuming the instructions are broken.

Revisit your instructions when your role changes, and roughly every few months regardless, because model updates shift how the boxes are read. The clearest signal that something belongs in the box is behavioral. The moment you catch yourself typing the same context into a fresh chat for the third time, stop. That repetition is the tax announcing itself. Move the line into Custom Instructions, and you never type it again. Treat the setup as a living thing you refine week by week rather than a one-time job, and you will watch your output quality make steady progress.

Custom Instructions fix who you are and how you always want ChatGPT to respond, within the limitations of what any model can do. The prompt library fixes what you need in the moment. Set the first once, keep the second close, and the generic-answer problem mostly disappears. What is left is the small share of work that genuinely needs a sharper prompt, and now you know which layer that belongs in.

Final Thoughts: Stop Prompting Harder, Start Configuring Smarter

The people who get the most out of ChatGPT are not the ones with the cleverest prompts. They are the ones who stopped treating every conversation as a fresh start. Once you move your durable context into Custom Instructions, tune your personality preset, and keep a short library of structural prompts for the tasks that change day to day, the tool stops fighting you. You spend your energy on the work instead of on re-explaining yourself, and the quality of what comes back climbs without any extra effort on your part. That is the whole shift, from prompting harder to configuring smarter, and it takes an afternoon to set up and pays off for as long as you use the tool.

The teams pulling furthest ahead are going one step past configuration into automation, wiring these same models into agents and workflows that run without a human typing each prompt. If that is where you want to go, Leland can help you get there with people who do this for a living.

• Go deeper with the Leland AI Builder Program, a live cohort that turns knowledge workers into people who ship real AI agents and workflows in a matter of weeks.
• Work one-on-one with an AI automation and agents coach who can tailor a setup to your exact role and stack.
• Or start free by joining an upcoming AI automation and agents event to learn live from operators before you commit to anything.

See also: Top 10 AI Consultants and Experts

Read next:

• The 5 Best AI Coding Agents: Pros & Cons, Reviews, & Which is Best for You
• The 5 Best AI Tools & Agents for Business: Reviewed & Ranked
• The 5 Best AI Tools & Agents for Developers: Reviewed & Ranked
• The 5 Best AI Voice Agents (By Type & Function)
• The 5 Best AI Newsletters to Subscribe to
• The 5 Best AI Tools & Agents for Finance: Reviewed & Ranked
• The 5 Best AI Tools & Agents for Video Editing: Reviewed & Ranked
• The 5 Best AI Personal Assistants: Reviewed & Ranked
• Agentic AI vs. AI Agents: Differences & What You Need to Know

FAQs

Do my custom instructions carry over if I use ChatGPT on my phone and laptop, or do I have to set them up again on each device?

• Your Custom Instructions are tied to your OpenAI account, not the device, so once you set them, they apply everywhere you log in, including web, the desktop app, and iOS and Android. The same holds for your personality preset and Memory. The one thing that does not travel is a specific conversation's context, since that lives in the individual chat. If you set instructions on your laptop and they do not appear on your phone, you are almost always looking at a different account or a stale app that needs a refresh.

Can other people see my custom instructions if I share a chat link or hand my screen to a coworker?

• A shared chat link shows the conversation itself, not the Custom Instructions behind it, so your background details and rules stay private even when you send someone a link to an answer. That said, the output was shaped by those instructions, so if you told ChatGPT you are job hunting or wrote in sensitive personal context, that framing can still show up in the response you share. Screen sharing is the real exposure risk, since anyone watching can open your Settings. Keep genuinely private details out of the boxes and lean on per-message context for anything you would not want a colleague to read.

Will using custom instructions use up my messages faster or cost me more on the paid plan?

• No. Custom Instructions do not count against your message limits or add a separate charge on any consumer plan. They are added as context behind the scenes, so from your side, a message is a message whether the boxes are full or empty. The only real cost is a small amount of the model's context budget, which matters for very long conversations rather than for your bill. This is different from the API, where every token, including system context, is billed, but that is a developer concern, not something that affects you in the ChatGPT app.

Is it a bad idea to just copy someone else's custom instructions I found online instead of writing my own?

• Copying is a fine starting point and a bad ending point. A set you find online is calibrated to someone else's role, audience, and quirks, so pasting it whole gives you their defaults, not yours. The safe move is to keep the structural lines that are role-neutral, such as anti-flattery rules or format preferences, and replace every specific detail with your own before you save. Watch especially for lines that quietly assume a job or a writing style that is not yours, because those will steer output in a direction you did not ask for and will not immediately notice.

If I turn off Memory, do my custom instructions stop working too?

• No, they are separate systems. Memory and Custom Instructions sit in the same Personalization area and are easy to confuse, but turning off Memory only stops ChatGPT from drawing on saved facts and past-chat patterns. Your Custom Instructions keep applying to every new chat exactly as written. Some people actually prefer this combination, which is instructions on for reliable, controlled context and Memory off to avoid stale details creeping in. If you want ChatGPT to know your role and rules but not to quietly accumulate assumptions about you over time, that is the setup to use.