AI Custom Instructions: The Setup Most Beginners Skip

Every AI conversation you start is a blank slate. The AI knows nothing about you. Not your job, not your writing style, not whether you prefer bullet points or paragraphs. So you explain yourself. Again. And again. And again.

Most advice about AI custom instructions focuses on prompt tricks and clever wording. But the single biggest improvement I made to my AI results had nothing to do with how I wrote prompts. It was what I set up before I started typing.

Custom instructions are the persistent preferences you configure once, and every conversation after that inherits them automatically. Think of them as permanent context. You tell the AI who you are, how you work, and what you care about. Then it remembers. Every time.

Here’s what surprised me: the difference wasn’t subtle. It was like going from talking to a stranger to talking to someone who already knows the brief.

What AI Custom Instructions Actually Do

Every major AI tool now offers some version of this feature. The names are different, but the concept is identical.

ChatGPT calls them Custom Instructions. You fill in two fields: what ChatGPT should know about you, and how it should respond. These apply to every new conversation.

Claude uses Projects for detailed setups (Pro plan) and User Preferences in settings for instructions that follow you everywhere. You can also upload reference documents that Claude reads alongside your instructions. With the latest Claude Sonnet 4.6 model, improved instruction following makes custom instructions even more effective.

Gemini has Gems, which are pre-configured assistants you build for specific tasks. Each Gem carries its own instructions and personality.

The mechanics vary. The outcome is the same: you stop repeating yourself, and the AI starts responding like it actually knows who it’s talking to.

Why Most People Skip This (and Why It Costs Them)

I skipped custom instructions for months. The reasoning felt logical: I’m already good at writing prompts, so why bother with setup? Just be more specific in the moment.

That reasoning is wrong.

Here’s what happens without custom instructions. You ask for help writing an email and get something that sounds like a corporate press release. You ask for code and get verbose explanations you didn’t need. You ask for feedback and get three paragraphs of validation before a single useful suggestion.

The AI isn’t being unhelpful. It just doesn’t know you. So it defaults to the safest, most generic version of helpful it can manage.

If you’ve already read about how to give AI context, you know that specificity transforms results. Custom instructions are that same principle, made permanent. Instead of providing context every session, you provide it once and it sticks.

The Before and After That Convinced Me

Let me show you what changed. Same prompt, two completely different results.

The prompt:

Help me write a short LinkedIn post about a lesson I learned this week.

Without custom instructions, I got this:

Certainly! Here's a LinkedIn post for you:

🚀 This week I learned an important lesson about growth...

In today's fast-paced world, it's easy to get caught up in the
hustle. But sometimes the most powerful thing you can do is slow
down and reflect.

Here are 3 key takeaways:
1. Embrace failure as a learning opportunity
2. Seek feedback from diverse perspectives
3. Stay curious and never stop learning

What lessons have YOU learned this week? Drop them in the
comments below! 👇

#Growth #Leadership #LessonsLearned

Generic. Hashtag-stuffed. Could have been written by anyone about anything.

With custom instructions set up, the same prompt produced:

Spent the week trying to automate something that should have
taken an afternoon.

The tool worked. Eventually. But I spent more time configuring
the automation than I would have spent doing the task manually
for the next six months.

Not every problem needs a system. Sometimes the answer is just
doing the thing.

Night and day. The second version sounds like a real person because the AI knew I prefer direct language, short paragraphs, no emojis, and honest observations over polished corporate content.

The instructions made the difference. Not a better prompt.

Copy-Ready Custom Instructions You Can Use Today

Here’s where this gets practical. Below are starter templates you can copy, paste, and adjust. I’ve written them for real situations, not job titles.

The Everyday Starter

This works for anyone. Paste it into ChatGPT’s custom instructions, Claude’s user preferences, or use it as the foundation for a Gemini Gem.

About me:
I'm not a technical person. I use AI to help with writing,
planning, and figuring things out. I prefer clear, practical
answers over theoretical explanations. If something has steps,
walk me through them one at a time.

How to respond:
- Write in plain English. No jargon unless you explain it.
- Keep responses short unless I ask for more detail.
- Give me one good answer rather than five options.
- If you're unsure about something, say so. Don't guess.
- No emojis. No hashtags. No bullet points unless I ask.
- Use British English spelling and grammar.

The Writer

For anyone who uses AI to help draft content, emails, or social posts. If you want to go further, you can make AI match your writing style by building a reusable style prompt from your own writing samples.

About me:
I write content for [your audience, e.g. "a blog about
personal finance" or "internal team communications"]. My style
is conversational and direct. I prefer short sentences and
paragraphs. I avoid corporate language and buzzwords.

How to respond:
- Match my writing style: conversational, clear, no fluff.
- Never open with "Certainly!" or "Great question!" Just start.
- Write drafts I can edit, not finished pieces. Leave room for
  my voice.
- If I share a draft, give honest feedback. Tell me what's
  weak, not just what's good.
- Suggest stronger words when mine are vague.
- Use British English throughout.

The Problem Solver

For people who bring AI their real-world dilemmas.

About me:
I use AI to think through decisions and solve problems. I'm
not looking for generic advice. I want you to ask me clarifying
questions before jumping to solutions. I value trade-offs being
made explicit.

How to respond:
- Before answering, check if you have enough context. Ask if
  you don't.
- Present options with honest pros and cons. Don't recommend
  the safe choice by default.
- Be direct. If an idea is bad, say so and explain why.
- Keep your answers focused. One clear recommendation is better
  than a menu of possibilities.
- If my question is vague, push back and help me sharpen it.

The Learner

For people using AI to understand new topics.

About me:
I'm learning about [your topic, e.g. "how websites work" or
"investing basics"]. I'm a beginner. I learn best from
examples and analogies, not definitions and theory. Don't
assume I know technical terms.

How to respond:
- Explain concepts with everyday analogies first, then give
  the proper explanation.
- Use concrete examples. Show me what something looks like in
  practice.
- Check my understanding by asking me to explain it back
  (when I ask you to).
- Build on what I already know rather than starting from
  scratch every time.
- If a topic has a rabbit hole I should avoid as a beginner,
  tell me.

Three Things Most Guides Get Wrong About Custom Instructions

I’ve read a lot of custom instructions advice. Most of it overcomplicates things.

‘Use a framework’

You’ll find articles promoting acronyms like RODES or CRISP or any number of structured approaches. These are fine if you enjoy frameworks. But for most people, they add friction where there shouldn’t be any. Writing your custom instructions in plain language works better than slotting your preferences into someone else’s template. Say what you mean. The AI will understand.

‘Write separate instructions for every platform’

Some guides suggest maintaining different custom instructions for ChatGPT, Claude, and Gemini. Unless you genuinely use all three for different purposes, start with one. Get it right. Then copy it across if needed. The principles are the same everywhere.

‘Be as detailed as possible’

More is not always better. Custom instructions that run to 1,500 characters of dense specification can actually confuse the AI or get partially ignored. Start with five or six clear statements. Add more only when you notice specific problems in the responses you’re getting. Build up through use, not through imagination.

How to Refine Your Instructions Over Time

The first version of your custom instructions will not be perfect. Mine weren’t. They still aren’t, really.

The trick is to treat them as a living document. When something annoys you in an AI response, don’t just fix it in the conversation. Go back to your instructions and add a line that prevents it from happening again.

A few examples from my own iteration:

I kept getting responses that opened with filler phrases like ‘Absolutely!’ and ‘That’s a great point!’ So I added: ‘Never open with affirmations or filler. Start with substance.’

The AI kept giving me five options when I wanted one recommendation. So I added: ‘Give me your single best suggestion unless I ask for alternatives.’

Responses were too long for quick questions. So I added: ‘Match response length to question complexity. Simple question, short answer.’

Each tweak took ten seconds. Each one made every future conversation slightly better. That compounds.

When Custom Instructions Aren’t Enough

Custom instructions work brilliantly for your general preferences. But some work needs more than preferences. It needs project-specific context.

If you’re working on something substantial (a writing project, a business plan, a learning curriculum) you’ll want to set up a dedicated space. Claude’s Projects feature lets you upload documents and write detailed instructions that only apply within that project. ChatGPT’s custom GPTs serve a similar purpose.

If you use Claude Code for development work, the concept extends even further. The CLAUDE.md file in your project acts as custom instructions for your entire codebase. I’ve written about how that file can grow out of control if you’re not careful, but the principle is the same: give the AI context once, get better results every time.

For now, though, start with the basics. Get your general custom instructions right. That alone will change how useful AI feels on a daily basis.

Stop Explaining Yourself Every Session

The best prompt technique isn’t a technique at all. It’s setup.

Five minutes configuring your AI custom instructions will save you from repeating yourself in every conversation you have from now on. The templates above are ready to copy. Pick the one closest to how you use AI, paste it in, and adjust it to sound like you.

Then start a new conversation and notice the difference.