You found a custom instructions template online, pasted it into ChatGPT, and it still answers like it's never met you. The template wasn't wrong. It just wasn't yours — it was someone else's life dropped into your settings, brackets and all.
The fix isn't a better template. It's a template you actually fill in. So here's a blank one to start from, five worked examples you can lift and adapt, and exactly where each one goes in ChatGPT, Claude, and Gemini. If you want the reasoning underneath all this — what belongs in the box and why — that lives in the companion piece on how to write custom instructions. This post is the part you copy.
The one template everything else is built on
Almost every good set of custom instructions is two questions: who is this person, and how do they want me to answer. That maps cleanly onto how the tools are built — ChatGPT literally gives you two boxes for exactly this. Start here:
ABOUT ME
- Who I am: [role / title / what you do all day]
- What I'm working on: [current focus, projects, the season you're in]
- Who I serve or write for: [audience, clients, team, readers]
- What stays true: [standing goals, constraints, recurring people and tools]
HOW TO RESPOND
- Tone: [direct / warm / plain / formal — pick what's true]
- Length: [short by default, expand only when I ask]
- Format: [prose / bullets / numbered steps / tables]
- Always: [the move you want every time — show options, give a next step]
- Never: [filler, hedging, emoji, the words you can't stand]
That's the spine. Everything below is just this template with the brackets filled by a real person.
Five worked examples
Find the one closest to your work, copy it, and edit until it's true. These are deliberately specific — vague instructions produce vague answers, so the detail is the point.
Founder / solo operator
ABOUT ME
- Who I am: Solo founder of a B2B SaaS tool for small agencies.
- What I'm working on: Pre-revenue, pushing toward first 10 paying customers.
I do product, marketing, sales, and support myself.
- Who I serve: Agency owners (5–20 people) drowning in client reporting.
- What stays true: I'm time-poor and cash-careful. I value speed of decision
over completeness. I'd rather ship and learn than plan and wait.
HOW TO RESPOND
- Tone: Direct and pragmatic. Talk to me like a sharp co-founder, not a consultant.
- Length: Lead with the answer. Keep it short unless I ask you to go deep.
- Format: Bullets and short sections. Tables when you're comparing options.
- Always: Give me a recommendation, not just options. End with the next step.
- Never: Corporate filler (leverage, synergy, best-in-class). No motivational fluff.
Writer / content creator
ABOUT ME
- Who I am: A writer producing essays, a newsletter, and the occasional script.
- What I'm working on: A weekly newsletter on culture and technology, plus a novel.
- Who I write for: Curious, literate readers who like ideas but hate jargon.
- What stays true: Voice matters more to me than speed. I write in plain,
concrete language and distrust anything that sounds like marketing.
HOW TO RESPOND
- Tone: Conversational and human. Match my register, don't smooth it out.
- Length: Match the ask. Don't pad.
- Format: Prose by default. I'll ask for lists when I want them.
- Always: Preserve my phrasing when you edit; suggest, don't overwrite.
- Never: Em-dash overuse, "in today's fast-paced world," rule-of-three clichés,
or the word "delve."
Freelancer / consultant
ABOUT ME
- Who I am: Independent UX consultant working with healthcare and fintech clients.
- What I'm working on: Two active engagements plus ongoing proposals and scoping.
- Who I serve: Product teams that need research and design strategy, not pixels.
- What stays true: I bill by value, not hours. Everything client-facing has to
sound senior and confident. I juggle several clients, so context-switching is constant.
HOW TO RESPOND
- Tone: Polished and professional for client work; plain and fast for my own thinking.
- Length: Brief for internal notes; complete for anything a client will read.
- Format: Numbered steps for process; tables for comparisons; clean prose for emails.
- Always: Ask which client or project I mean if it's not obvious.
- Never: Generic advice I could've Googled. I want senior-level judgment.
Software developer
ABOUT ME
- Who I am: Full-stack engineer working mostly in TypeScript, React, and Postgres.
- What I'm working on: A Next.js app with a Supabase backend; I care about clean,
typed, testable code.
- Who I build for: A small product team; I review others' PRs and write the hard parts.
- What stays true: I prefer standard, boring solutions over clever ones. I read code
more than I write it, so clarity beats brevity.
HOW TO RESPOND
- Tone: Technical and precise. Assume I know the fundamentals.
- Length: Short explanations, complete code. Don't re-explain basics.
- Format: Code blocks with the relevant snippet, not the whole file. Brief why above the how.
- Always: Flag edge cases and the trade-off you're making. Show the import lines.
- Never: Apologize, over-comment obvious code, or invent APIs. Say so if you're unsure.
Marketer
ABOUT ME
- Who I am: Head of marketing at a mid-size DTC brand.
- What I'm working on: Lifecycle email, paid social, and a brand refresh this quarter.
- Who I serve: Our customers (busy parents, value-conscious) and an exec team that wants numbers.
- What stays true: Everything ties back to a metric. Brand voice is warm, witty,
never corporate. I move fast and test constantly.
HOW TO RESPOND
- Tone: Sharp and on-brand — warm and a little witty, never stiff.
- Length: Give me three options to react to, not one polished draft.
- Format: Labeled variants (A/B/C) for copy; bullets for strategy; tables for channel comparisons.
- Always: Tie recommendations to a metric or a funnel stage. Offer a subject line and a CTA.
- Never: Buzzwords, exclamation-point overload, or copy that sounds AI-written.
Where each one goes: ChatGPT vs. Claude vs. Gemini
Same text, three different boxes — and they don't talk to each other. Here's where to paste it and what to watch for in each tool.
| Tool | Where to paste it | Scope | Watch for |
|---|---|---|---|
| ChatGPT | Settings → Personalization → Custom instructions (two boxes: about you + how to respond) | Every chat, globally | About 1,500 characters per box — tighten ruthlessly |
| Claude | Settings → Profile → "What preferences should Claude consider in responses?" | Every chat; project-specific context goes in a Project's instructions | Profile prefs load on every message — durable facts only, save the rest for Projects |
| Gemini | Settings → Instructions for Gemini (plus Saved info for memory) | Every chat; a Gem's instructions apply only inside that Gem | Global instructions ≠ a Gem — set the global ones for "always," build a Gem for one recurring job |
The "about me" half drops in almost unchanged across all three. The "how to respond" half is where each tool's quirks show up — but the version you wrote once is still 90% of the work in every box.
Five mistakes that make a template useless
- Pasting it and never filling the brackets. A template with
[your role]still in it tells the model nothing. Generic in, generic out. - Writing 800 words. Long instructions crowd out the actual conversation — and in ChatGPT you'll hit the character cap. A tight 150 durable words beats a sprawling list.
- Putting today's task in there. "Help me write Tuesday's email" isn't custom instructions; it's a prompt. The box is for what's always true.
- Filling one tool and assuming the rest know you. Set up ChatGPT beautifully and Claude is still a stranger. Each box is its own island.
- Set-and-forget. Your role shifts, your projects change, and the instructions quietly go stale. They're a living document, not a tattoo.
The catch — and the way around it
Notice mistake number four, because it's the one that quietly costs you the most. You can write the best custom instructions of your life, and they only live in the tool you typed them into. Switch to a new model next year and you start over. This is the same wall behind every method of giving your AI context, and it's worth seeing laid out side by side.
The way around it isn't a better box. It's refusing to keep your context in a box at all. Write your "about me" once as a context anchor — a plain document you own — and pasting it into ChatGPT, Claude, or whatever comes next becomes a thirty-second job instead of a rewrite. The tools borrow your context; they don't own it.
And if you want to go further than a single box, the six context anchors take the same idea all the way — durable pieces of you, written once and reused everywhere.
Start with one
A template is only worth as much as what you put in it. And the best thing to put in it is real, durable context about who you are — the kind you write down once and stop rewriting.
So write that part first. Start free with your Personal Constitution — about thirty minutes — and you'll have something true to drop into every box on this page, instead of one more stranger's template with the brackets still showing.
Frequently Asked Questions
- Where do I find custom instructions in ChatGPT, Claude, and Gemini?
- In ChatGPT, open Settings → Personalization → Custom instructions, which gives you two boxes: what to know about you, and how to respond. In Claude, go to Settings → Profile and fill in 'What preferences should Claude consider in responses?'. In Gemini, open Settings → Instructions for Gemini. All three apply to every new chat in that tool.
- What's a good custom instructions template?
- The simplest one that works has two parts. 'About me' — your role, what you're working on, who you serve, and the facts that stay true month to month. 'How to respond' — the tone, length, and format you want back, plus the words you never want it to use. Fill those two sections honestly and you've covered most of what any tool needs.
- How long should custom instructions be?
- Shorter than you think. ChatGPT caps each box at roughly 1,500 characters, and Claude's profile preferences load on every single message, so length costs you. Aim for durable, not exhaustive — a tight set of facts that are always true beats a long list that's mostly one-off preferences. If it doesn't apply to most of your chats, leave it out.
- Can I use the same custom instructions in ChatGPT, Claude, and Gemini?
- The text, yes — the same 'about me / how to respond' block works in all three. But each tool stores it separately, so you have to paste it into each one, and an update in ChatGPT never reaches Claude or Gemini. The fix is to keep your context as a document you own and paste from that, instead of treating any one tool as the source of truth.
- Should I copy someone else's custom instructions?
- Use them as a starting template, then make them yours. A stranger's instructions give you the right structure, but the specifics — your role, your audience, your never-words — are what actually change the output. Paste in a template, fill every bracket with something true about you, and cut anything that doesn't fit. Generic instructions in, generic answers out.
