How to Use ChatGPT Custom Instructions (2026 Guide)

Custom instructions are the setting that stops you from re-typing “answer concisely, no preamble” at the top of every ChatGPT chat. You fill two boxes once, and the rules stick to every new conversation. We set this up on a Free account in May 2026 and watched the very next chat drop its usual “Certainly! Here’s a breakdown…” opener entirely.

#Where Are ChatGPT Custom Instructions in 2026?

Open ChatGPT, click your account name in the bottom-left, and pick Settings. Select the Personalization tab, then open Custom instructions. That’s the whole path on the web app, and the mobile apps match it almost exactly.

The mobile route is nearly the same. Tap your profile picture, then Personalization, then Custom instructions on the iPhone or Android app. The two text boxes look identical to the desktop version, and because the setting lives on your account rather than the app, anything you save shows up on every device you sign into.

According to Tom’s Guide’s walkthrough of the feature, custom instructions now sit under the Personalization tab and are part of the free plan. No paid account needed.

The rules follow your account, not the app you opened it from. If you launch ChatGPT from an Action Button shortcut or through Siri, the same instructions load, though the Siri route also depends on your device meeting the Apple Intelligence requirements for that integration to function on older hardware.

Check one thing before you leave.

The Enable for new chats toggle has to be on, or you can save all the text you want and nothing changes, because the instructions never load. Turn it on, then save.

#What to Put in Each Box

The first box asks “What would you like ChatGPT to know about you to provide better responses?” This is context: your job, your skill level, the tools you use, the audience you usually write for. The second box asks how you want ChatGPT to respond, which covers tone, length, and format.

Keep each box short.

OpenAI’s custom instructions documentation states that each long-form field holds 1500 characters, but you almost never need that much. A few precise lines beat a paragraph of vague preferences, because ChatGPT can’t tell which sentence in a wall of text is the one that matters.

You don’t have to fill both boxes. Leaving the “about you” box empty and using only the response-style box is a clean way to start, and we’d actually recommend it: write three lines about response style first, live with it for a few days, then add an “about you” block only once you spot what ChatGPT keeps getting wrong about your context.

The response box does most of the work. Telling ChatGPT “be direct, skip the recap, answer before the reasoning” reshapes nearly every reply. The “about you” box matters more for niche cases, like flagging that you write iOS code and want Swift examples instead of pseudocode.

#Custom Instructions Examples You Can Copy

We tested all four of these blocks on a Free account before listing them here. Paste one into the response box, tweak it, and save.

Concise, no fluff:

Be direct and skip preamble. Answer the question before explaining the reasoning. Keep most replies under three short paragraphs unless I ask for depth, and never open with “Certainly” or “Great question.”

Coding helper:

I write TypeScript and Swift. Default to working code over prose, with comments only where the logic isn’t obvious. Flag deprecated APIs, and when a question has more than one valid approach, show the simplest first, then note the trade-off in a sentence.

Writing and editing:

Edit for clarity at a grade-9 reading level. Cut filler and hedging, but keep my voice and don’t replace my word choices wholesale. When you suggest a rewrite, show the before and after so I can see what changed.

A named role:

Act as a skeptical product reviewer named Mara. Push back on weak claims, ask for evidence, and don’t agree with me just to be agreeable. Stay in this role unless I say “drop the persona.”

These work the same whether you’re on ChatGPT or a rival assistant. If you’ve set up profiles in Claude, the structure transfers: state the role, the audience, and the format. In our testing, a 1500-character instruction wall made it harder to tell which rule changed the behavior, so we trimmed ours to four lines and the format held fine.

#How to Edit or Remove Your Instructions

Changing them later takes the same path: Settings, then Personalization, then Custom instructions.

Edit the text, save, and the new version takes over from your next chat onward. To switch the feature off entirely, flip the Enable for new chats toggle back off, or just clear both boxes.

One catch surprises people. Editing or deleting your instructions doesn’t rewrite chats you’ve already had, because the old text was baked in when those conversations started. The OpenAI documentation notes that updates apply only to future conversations, and removing instructions from old chats means clearing those chats from your history. So if a months-old conversation still answers in the style you’ve since dropped, that’s expected, not broken.

#Custom Instructions vs Memory vs Projects

People mix these three up constantly, so here’s the clean split.

Custom instructions are fixed rules you write by hand. They don’t change unless you edit them, and they apply account-wide to every new chat. Think of them as a standing order.

Memory is different.

According to OpenAI’s Memory documentation, it works through 2 parts: saved memories you ask it to keep, and details it pulls from your past chats automatically. You can view and delete memories, but you don’t author them line by line the way you do with custom instructions. Memory learns; custom instructions are told.

Projects add a third layer.

Inside a workspace, you can write project-only instructions that take priority over your global custom instructions for that workspace. Our ChatGPT Projects guide covers the full setup, but the short version is that a Project’s instructions win locally while your global ones keep running everywhere else.

If you also use Google’s assistant, the closest equivalent to a packaged, instruction-driven workspace is Gemini Gems, which bundles a role and reference files the way a Project does. The mental model carries across tools: global preferences at the account level, scoped overrides at the workspace level.

#Why Aren’t My Custom Instructions Working?

Start with the Enable for new chats toggle. This is the number one cause we see. If it’s off, your text is saved but inert.

Next, remember that instructions only affect new chats. An old conversation you opened before saving won’t suddenly obey the rules, because the instructions loaded at the start of a chat and don’t reach back in time. Start a fresh chat to test any change.

Still ignored? Check for a conflict.

When you’re inside a Project, that Project’s instructions override your global ones, so a “be concise” global rule loses to a Project rule that says “explain in detail.” Long, contradictory instructions also confuse the model. Telling it to “be brief but thorough and exhaustive” gives it nothing clear to follow.

Finally, the model isn’t perfect at obeying every line on every turn. A single ignored instruction isn’t a settings failure. If it keeps dropping the same rule, shorten it, move it to the top of the box, and phrase it as a hard constraint.

#Bottom Line

Start small.

Put two lines in the response box (“be direct, skip preamble, answer before explaining”) and leave the “about you” box empty for the first week. Once you notice what ChatGPT keeps getting wrong about your context, add a short “about you” block that fixes exactly that.

Don’t paste a 1500-character wall on day one.

You won’t be able to tell which line is doing the work, and trimming a bloated block later is harder than adding to a lean one. If a rule isn’t sticking, check the toggle first, then start a new chat to test.

#Frequently Asked Questions