The fastest way to make ChatGPT or Claude actually sound like you isn't a better prompt or a different tool. It's the field most people have never touched. Custom instructions tell the model who you are, what you do, what you want it to sound like, and what you don't want it to do, once, so every chat after that already knows.
Filling it in takes about five minutes and changes everything the AI produces from that point on. The five things worth putting in are:
- Your role and what you do
- The tone you want it to use
- A ban list of words you never want it to write
- Context about your work and audience
- Your output format preferences
The rest of this post walks through each one with specific examples and the common mistakes worth avoiding.
What custom instructions actually are
Custom instructions are a settings field built into ChatGPT and Claude. You write them once, and they become standing context for every conversation after that. The model reads them at the start of every chat and uses them to shape its output, the same way a new colleague would read a one-page brief about your team before joining a meeting.
In ChatGPT, you find them under Settings, Personalization, Custom Instructions. There are two fields, each capped at 1,500 characters. The first asks what you'd like the model to know about you, and the second asks how you'd like it to respond.
In Claude, the equivalent lives at the Project level. You create a Project for a piece of recurring work, write the instructions in the Project description, and every conversation inside that Project inherits them. There's no published character limit, but Project instructions consume context tokens on every turn, so brevity is sensible.
The reason this matters is that without custom instructions, the model is starting from zero every time. It doesn't know your job, your tone, your audience, or the words you'd never use, so it defaults to a kind of generic LinkedIn voice that's recognisable as AI within about three sentences. Custom instructions are the difference between an output that needs rewriting and one you can actually send.
The five things worth putting in
Custom instructions work best when they're specific, short, and structured around what the model needs to know to be useful. Five categories cover most of it.
Who you are and what you do. Your role, your industry, who you work with, the kind of decisions you make. Not your CV, just enough context that the model knows what kind of output is appropriate. "I'm a marketing manager at a B2B SaaS company. I write internal updates, briefs for the design team, and emails to product managers and stakeholders." That's plenty.
How you want it to sound. The tone you actually use, in plain words. "Direct, warm, mid-length sentences. British English. No corporate jargon. I'd rather sound a bit informal than a bit stiff." The model is more responsive to descriptive tone instructions than to abstract ones, so "sound like a confident manager talking to a peer" works better than "sound professional."
The ban list. This is the highest-leverage line in the whole field: words and phrases you never want the model to use, listed plainly. "Never use: delve, tapestry, transformative, streamline, in today's fast-paced world, I hope this finds you well, navigate, leverage, robust, comprehensive, in conclusion." The ban list is what stops the recognisable AI vocabulary slipping into outputs, and is the single fastest way to make the writing sound like yours, not the model's.
Context about your typical work. What you do most often, who reads it, what success looks like. "Most of what I write is read by senior stakeholders who skim. They want the headline first and the detail second. Bullet points are fine. Long preambles are not." This gives the model a sense of format before you've asked anything, which means fewer rounds of editing.
Output format preferences. How you want responses structured by default. "Lead with the answer. Show working only if I ask. Keep responses short unless I say otherwise. No emoji. No headers unless the response is genuinely long." This is the line that stops every output from coming back as a numbered list with bold headings.
Examples that work
A marketing manager: "I'm a marketing manager at a B2B SaaS company writing internal briefs, external campaigns, and stakeholder updates. I use British English, prefer direct mid-length sentences, and would rather sound informal than corporate. Never use: delve, leverage, robust, transformative, navigate, in today's fast-paced world. Lead with the answer. Long preambles waste my time."
An ops lead: "I run operations at a 200-person company. Most of what I write goes to senior leadership and the board. Tone should be calm, precise, and structured. Reports lead with the headline, not the context. I work in a regulated industry, so don't invent statistics or sources. If you're not sure about a number, say so."
Someone earlier in their career: "I'm five years into my career in marketing and trying to come across as more senior than I sound naturally. Help me write more confidently and less apologetically. Never start emails with 'Sorry to bother you' or 'Just checking in'. Cut hedging language. Don't use exclamation marks. Match the tone of someone who's been doing the job ten years longer than I have."
The pattern across all three is specific role, specific tone, specific ban list, specific output rule. Vague custom instructions produce vague outputs, which is why most people give up on them after one attempt.
What people get wrong
- Writing it like a personality profile. "I'm friendly and curious and I love learning new things" tells the model nothing useful. A working brief tells it what to do.
- Leaving the ban list blank. Without one, the recognisable AI vocabulary will leak into every output, no matter how good the rest of the instructions are. The ban list is doing about half the work.
- Treating it as a one-time exercise. A good custom instructions field gets updated when things change, when you start a new project, when you notice the model defaulting to something annoying. Five minutes every couple of months keeps it useful.
- Skipping the audience. "Write for senior stakeholders" produces different output from "write for the design team," which produces different output again from "write for someone reading this on their phone in a meeting." The audience shapes the format, and the model can't guess.