What is an instructions prompt?

Imagine you have the world’s smartest intern working for you. Let’s call him Chad.

Chad has passed every test with flying colors. He aced the MCAT, got a perfect score on the bar exam, and solved some of mathematics’ hardest problems for fun.

However, that was a long time ago. He still remembers most of what he studied, but he relies on a quick internet search to double-check his work.

You’re still skeptical, so you decide to test him. His first assignment is simple: draft a reply to an email from your manager.

New chat

1. You

   Draft a reply to my manager. She wants to move our Thursday check-in to Friday morning. Friday works, but not before 10.

2. Chad

   Subject: Re: Rescheduling Our Thursday Check-In

   Dear [Manager’s Name],

   I hope this email finds you well. Thank you so much for reaching out regarding our check-in.

   I would be more than happy to move our meeting to Friday morning. However, I wanted to kindly mention that I have a prior commitment until 10:00 AM. Would a time after 10:00 AM work with your schedule?

   Please don’t hesitate to let me know if there is anything else I can do. I truly appreciate your flexibility!

   Warm regards,
   [Your Name]

It’s perfect. Too perfect. Like an AI wrote it. You give him some feedback.

New chat

1. You

   Make it sound like a human wrote it.

2. Chad

   Hi [Manager’s Name],

   Thanks for the heads-up! Friday morning works great for me. I’ve got something until 10, but anytime after that is perfect.

   Let me know what works best for you!

   Cheers,
   [Your Name]

3. You

   Make it less formal and more concise.

4. Chad

   Hey [Manager’s Name]! No worries at all. Friday works, anytime after 10. Let me know!

   Cheers,
   [Your Name]

Better. But still not quite there. You would never say some of the things he wrote. You have never signed an email “Cheers” in your life, and you don’t greet your boss with “Hey!” And he still doesn’t know her name, or yours.

How can this be the smartest intern in the world when he can’t even draft a simple email reply?

It’s because you haven’t given him enough context.

Every time you start a new chat, there is zero context. Without context, Chad does the best he can with what he has in front of him.

This is where the instructions prompt comes in.

How context works

Context is everything Chad can see at the moment he writes. Picture it as his desk.

What Chad studied stays in his head: grammar, medicine, the law, and how millions of emails usually go. What’s on his desk is your situation: your message, the email you’re answering, and any file you hand him. He writes from both.

In that first chat, the desk held one line from you. So Chad filled in everything else from his head, with the most likely email from the most likely sender. That’s why it sounded like AI. An email written for anyone sounds like no one.

Then the desk gets cleared. Every new chat starts empty. Chad doesn’t remember yesterday’s feedback, your manager’s name, or that you never write “Cheers.” Each new chat is his first day again.

A new chat

In his head

• What he studiedGrammar, facts, how most emails go

On his desk

• Your instructionsNothing here
• Your filesNothing here
• Your messageReply to my manager: Friday after 10 works

He writes

The most likely email, from anyone.

A chat in your project

In his head

• What he studiedGrammar, facts, how most emails go

On his desk

• Your instructionsWho you are, how you sound, what to avoid
• Your filesThree emails you wrote
• Your messageReply to my manager: Friday after 10 works

He writes

Your email.

What an instructions prompt is

An instructions prompt is a note you write once. The assistant reads it at the start of every chat, before your first message. Claude and ChatGPT call it instructions, custom instructions, or project instructions, and developers call it a system prompt. It’s the same idea.

Think about how a good manager starts a new hire. Before the first assignment, they sit down together for a morning: here’s who we are, here’s who you’ll work with, here’s how we talk to people, and here’s what we never do. The instructions prompt is that morning, written down. Chad reads it before every chat, so you only give the talk once.

The instructions aren’t the assignment. A brief tells Chad about one job. The instructions tell him about you, and they hold for every job.

	The brief	The instructions
Covers	One job	Every job
You write it	Each time	Once, then keep it current
Lives in	The chat	Your settings or a project
Sounds like	“Reply to my manager: Friday after 10 works.”	“Keep emails short. No exclamation marks.”

What goes in it

Five parts, written the way you would brief a new hire on the first morning. Say you’re Jordan, an operations manager at a beauty supply company, and your manager is Dana. Here’s what Jordan wrote:

1. Who you are. Your name, your role, and what your company does.

   “I’m Jordan, operations manager at a beauty supply company. We sell nail and spa supplies to salon owners.”

2. Who you write to. The people you write to most, and what they need from you.

   “My manager is Dana. She reads email on her phone, so put the answer first.”

3. How you sound. A few rules for your voice.

   “Write short and plain, with no exclamation marks. Sign off with ‘Jordan’ and nothing else.”

4. What to avoid. The words and habits you always cut from a draft.

   “Never write ‘Cheers’ or ‘No worries at all.’ Don’t end with an offer to help.”

5. When to ask. What to do when something is missing.

   “If you need a name, a date, or a fact I haven’t given you, ask me.”

Keep the whole thing to about a page. Chad reads all of it before every chat, and a long packet buries the lines that matter.

Jordan saved the instructions in a project and added three emails as files:

ProjectOnly you

Email replies

Read at the start of every chat in this project

Instructions

I’m Jordan, operations manager at a beauty supply company. We sell nail and spa supplies to salon owners. My manager is Dana. She reads email on her phone, so put the answer first. Write short and plain, with no exclamation marks, and sign off with “Jordan.” Never write “Cheers” or “No worries at all.” If you need a name, a date, or a fact I haven’t given you, ask me.

Files

Reply to Dana.pdfWeekly update.pdfNote to a salon.pdf

Start a chat in Email replies…

The same request, with instructions

Jordan opens a new chat inside the project and types the same line as before.

Email repliesProject

1. You

   Draft a reply to my manager. She wants to move our Thursday check-in to Friday morning. Friday works, but not before 10.

2. Chad

   Hi Dana,

   Friday works, but I’m booked until 10. Any time after that is fine.

   Jordan

That’s the email Jordan would have written. The request didn’t change, and Chad didn’t get any smarter. The only thing that changed is what was on his desk.

Set up yours

Claude and ChatGPT each have two places for instructions:

• Your account settings, for what’s true in every chat: who you are and how you sound. Keep this part short.
• A project, for one kind of work. A project is Chad’s own office. The instructions are pinned to the wall, the files sit in the cabinet, and every chat you start inside it starts with both.

Start with one project for the job you do most. If that’s email, call it “Email replies.”

Fill in the template

1. Download the instructions template.
2. Open it in any text editor. Notepad, TextEdit, and Word all work.
3. Replace each line in square brackets with your own words.
4. Save it under your own name, for example “Jordan’s instructions.”

Keep the file. It’s your master copy: when the instructions change, change the file first, then paste it in again. The same file works in Claude, ChatGPT, or any other assistant.

In Claude

1. Open Projects, and click New Project.
2. Give the project a name, such as “Email replies.” Claude doesn’t read the name or the description.
3. Click Set project instructions, paste in your filled-in template, and click Save instructions.
4. Click +, and add two or three emails you wrote.
5. Start a new chat from the project’s page.

For instructions in every chat, open Settings, then Your Claude, and fill in Instructions with the first and third parts: who you are and how you sound.

In ChatGPT

1. In the sidebar, click New project, and give it a name.
2. Open the project’s menu, and click Project settings.
3. Paste your filled-in template into the instructions, and save.
4. Add two or three emails you wrote as files in the project.
5. Start a new chat inside the project.

For instructions in every chat, open Settings, then Personalization, and fill in Custom instructions. Inside a project, the project’s instructions take priority over your custom instructions.

Test it

• Open a new chat inside the project.
• Send a request you’ve sent before, word for word.
• Compare the two drafts.
• Find one line you would still change, and add a rule for it to the instructions.

Keep it current

Onboarding doesn’t end after the first morning. When you correct Chad on the same thing twice, that’s a rule. Add it to your file, then paste the file into the project again.

Depending on your plan, you can share a project with your team. Then everyone’s Chad starts each chat with the same packet, and one change to the instructions reaches everyone.

Chad was always the smartest intern in the world. He didn’t know who he was working for. Now he does.