How to customize ChatGPT so it actually fits your work

If ChatGPT's answers feel flat, the usual reason is that it knows nothing about you. OpenAI's own 2025 usage study, analyzing 1.5 million conversations, found non-work messages made up more than 70% of all usage and that three use cases, practical guidance, writing, and seeking information, account for around 78% of messages. In other words, most people run it casually and stock, and the bulk of real work is exactly the writing and guidance that customization sharpens most. When you customize ChatGPT, you stop re-explaining who you are and what good looks like at the start of every chat, and it starts answering like a tool that knows your role instead of a stranger. The whole setup takes about ten minutes and it pays off on every future conversation, which is why it's worth doing once, deliberately. Where ChatGPT sits among the best AI tools for work hasn't changed; how much you get out of it has.

Everything lives in one place. Open Settings, then Personalization, and you'll find the two layers that do most of the work.

Custom instructions are two free-text boxes, each capped at 1,500 characters, and they're the single most useful thing you'll set. The first box, what ChatGPT should know about you, is where you put your role, your company or industry, who you write for, and what you do all day. The second, how it should respond, is where you set default length, format, tone, and what to skip, things like 'answer first, no preamble, no disclaimers.' Put evergreen guidance here and leave the changing facts to memory. One 2026 nuance worth knowing: there's now a personality preset on top of these boxes, Professional, Efficient, Friendly, and so on, and it can quietly fight what you typed, so make sure the preset and your text agree rather than contradict. These instructions are really just reusable prompting, the same moves covered in prompt engineering, saved once so you don't retype them every session.

Memory is the second layer, and it comes in two forms: saved memories, the facts you explicitly ask it to remember, and reference chat history, the patterns it picks up from past conversations. Seed it deliberately with the things you repeat constantly, 'I write for SMB owners,' 'we use this stack,' 'always cite sources,' using plain 'remember that' phrasing. In 2026 OpenAI added a more efficient memory architecture that synthesizes across conversations on its own and revises stale entries over time, and by its own internal measure factual recall improved from 67.9% to 82.8%. That helps, but it isn't perfect, so spend thirty seconds now and then in Manage memories deleting anything stale or wrong. And if you'd rather it remember nothing, memory is fully optional: toggle either layer off, clear individual entries, or use a temporary chat that bypasses it entirely. On a work account, you set that boundary.

If you do the same kind of work repeatedly, projects are where customization gets powerful. A project is a dedicated workspace that groups related chats, carries its own custom instructions layered on top of your global ones, and holds reference files, a brand guide, a style sample, a spec, that every chat in that project can draw on. Set up one project per ongoing job: your newsletter, your sales emails, a specific client. That way the context for each job stays put and doesn't leak between them, and you stop pasting the same background into every new chat. Projects reached free users in late 2025, so this isn't a paid-only move. For a team, shared projects are also how you keep output consistent across people, which is part of building AI into how a whole team works rather than leaving everyone to configure their own setup, role by role.

Two fair objections. The first: it ignores my instructions or forgets things. Real friction, and usually fixable, the personality preset overriding your typed instructions is the common culprit, so a two-minute audit to remove the contradiction solves most of it, and the 2026 memory update directly targets the forgetting. The second: it's not worth the effort. The math says otherwise, it's a ten-minute one-time setup against every future chat, and OpenAI's own data shows the heaviest use is writing and guidance, exactly what custom instructions improve. Frame it as a habit, not a project: set the basics today, then over the next week, when an answer still misses, tighten the instructions rather than starting over. That's the difference between a tool you fight and one that fits your work, and it's a skill any business can build deliberately across its people.