Instructions vs messages
Instructions are persistent, they set the role, tone, and rules for every reply, so you never need to repeat yourself.
Two kinds of text you give an AI
Every time you interact with a chat AI, you are giving it two different kinds of text, but most people treat them as the same thing.
A message is a one-off request: "Summarize this article." "Write a reply to this email." Each message is answered in isolation unless the conversation history carries some context forward.
Instructions are different. Instructions steer the AI's behavior across everything it does: who it is, how it writes, what it should and should not include. Set them once, and every reply that follows is shaped by them, you do not have to repeat yourself.
The difference matters because the AI cannot guess your preferences from a message alone. "Write an update email about our project delay" produces a generic draft. Add instructions that say you are a project manager writing for a senior leadership audience in a concise, direct style, and the same request produces something usable.
In the "with instructions" example below, the imagined user's instructions specified their role, audience, tone, and current project context, including the dates and the integration name. That context is what produces the specific reply; the AI did not invent those details.
Write an update email about our project delay.
Subject: Project Update, Timeline Adjustment Hi team, I wanted to reach out with an update on our project. We have run into some delays and the original timeline will need to be adjusted. We are working hard to resolve the situation and will keep everyone informed as things progress. Thank you for your patience and understanding. Best regards, [Your name]
Write an update email about our project delay.
Subject: Launch timeline update, 2 weeks Team, Launch moves from March 14 to March 28. Root cause: the payment-API integration took longer than scoped. No further slippage expected. What this means for you: the revised go/no-go is March 25. I'll send the updated schedule by Friday. , [Your name]
Toggle instructions, watch the reply change
Same question every time. Turn instruction rules on and off below and watch the AI's reply update live, that's instructions steering output.
Your question
"Write an update email about our project delay."
Your instructions
None set, the AI falls back to a generic, one-size-fits-all reply.
Where durable instructions live
Each major AI tool has its own place to store instructions that persist across conversations. You set them once; the tool sends them automatically with every message.
In ChatGPT, go to Settings → Personalization → Custom Instructions. Whatever you write there is sent to every chat. The ChatGPT track covers this in detail.
In the Claude app, create a Project and write instructions in the Project Instructions field. Every chat inside that project inherits them. The Claude App track walks you through setup.
In Claude Code, a file called CLAUDE.md in your project folder acts as persistent instructions. The model reads it at the start of every session. The Claude Code track explains the full pattern.
Armory's library gives you ready-made instruction sets for common roles, so you can start with something proven rather than writing from scratch.
Question 1 of 4
What is the key difference between an instruction and a message?