A flat-vector illustration on a deep navy field split by a single horizon line. Left of centre, a small sand-coloured sailboat sits on the surface. Below the waterline, its tapered keel and a wide ballast weight glow golden, far larger than the boat above, against soft washes of ocean blue and violet water.
All Resources

What Is a System Prompt? (And How It Differs From Custom Instructions)

By Chad Stamm · November 18, 2026 · 6 min read

Spend any time around AI tools and "system prompt" starts showing up everywhere. In product announcements. In leaked-prompt threads. In the settings page of an app you just installed.

It rarely comes with a definition. People use it the way they use "the algorithm," as if everyone already knows what it is and where it lives.

The idea is simpler than the name. Every AI conversation has a layer underneath the one you see, and that layer is doing a lot of the steering.

What is a system prompt?

A system prompt is a set of standing instructions an AI model reads before it reads your message. It tells the model who it is, who it's serving, what rules to follow, and how to answer. It applies to every turn of the conversation, not just the first one.

You type a question. The model doesn't start from your question. It starts from the system prompt, then reads your question in light of it.

That's why the same underlying model can feel like two different products. Put it inside a customer-support app and it's polite, narrow, and on-script. Put it inside a coding tool and it's terse and technical. Same engine. Different instructions underneath.

Anthropic's own prompting guide puts the power of it plainly: "Setting a role in the system prompt focuses Claude's behavior and tone for your use case. Even a single sentence makes a difference."

Where does the system prompt sit in a conversation?

Think of a conversation as layers, read from the top down.

The first layer belongs to whoever built the app. Anthropic, for example, publishes the system prompt it uses in Claude's web and mobile apps, which it says provides "up-to-date information, such as the current date, to Claude at the start of every conversation." The same page notes that these updates "do not apply to the Claude API." Developers building on the API write their own.

The next layer is the one a regular user can reach: custom instructions, project instructions, a Gem, a CLAUDE.md file. Your standing context.

The bottom layer is the message you type right now.

The layers above generally carry more weight than the ones below. OpenAI's API documentation says it directly: developer messages "are instructions provided by the application developer, prioritized ahead of user messages." OpenAI now calls that slot a developer message rather than a system message. Google's Gemini API calls it a system instruction. Anthropic's API calls it the system parameter. Three names, one job.

System prompt vs custom instructions vs a regular prompt

These three get confused constantly, so here they are side by side.

System prompt Custom instructions Regular prompt
Who writes it The app's builder, or a developer using the API You, inside a consumer app You, every turn
When it applies Every turn of every conversation in that app Every new chat in that one tool This message only
Can you see it Usually not Yes, and edit it Yes
What it's for Role, rules, format, guardrails Who you are and how you want answers The task in front of you
Where it lives Inside the app or your code Inside one vendor's settings The chat box

The middle column is the interesting one. Custom instructions, Claude Projects, and Gemini Gems are the consumer versions of the same idea. They let you write standing instructions without touching any code. The app still runs its own instructions alongside yours. You just get to edit your part.

Can you write your own system prompt?

Yes, and you probably already have something close to one.

If you use the API, you write it directly: the system parameter for Claude, a developer message or the instructions parameter for OpenAI, system_instruction for Gemini. Whatever you put there frames every response your app gets back.

If you don't write code, your system prompt is spread across settings. The custom-instructions box in ChatGPT. The instructions on a Claude Project. The description of a Gem. If you've started using Claude Code, it's the CLAUDE.md file that loads at the start of every session.

The thing is, most people fill those boxes with a line about tone and stop there. "Be concise. Be friendly." That's etiquette. It tells the model how to talk, and nothing about who it's talking to.

What goes in a good system prompt?

Five parts cover almost every use.

  • A role. One sentence on what the model is for.
  • Context it can't guess. Who it's working for, what that person does, how they decide, and what's current.
  • Rules you could check. "Keep client emails under 150 words" is a rule. "Be clear" is a wish.
  • A format. How long, how structured, what to always include.
  • One example. A short answer you'd accept, so the model has a target rather than a description.

Here's what that might look like for a made-up independent consultant:

# Role
You are a writing and planning assistant for an independent
brand consultant who works with small food and beverage companies.

# About them
- Decides on clarity first, then speed, then polish
- Two client rebrands due before December 15

# Rules
- Keep emails under 150 words
- Never invent pricing. Ask.
- Plain language. No jargon.

# Format
Short paragraphs. Lead with the answer, then the reasoning.

Notice what does most of the work. The role is one line. The context is where the answers stop sounding generic, because that's the part the model has no other way to learn. It's the same reason context engineering has become its own discipline: the instructions matter, but what you tell the model about the situation matters more.

What a system prompt can't do

Three limits worth knowing before you lean on one.

Priority isn't a guarantee. Providers design models to put standing instructions ahead of a user's message, and that is a strong lean, nothing more. Anything that truly must never happen belongs in the app's permissions and settings, not just in a sentence of instructions. Giving an AI agent context covers that second layer.

It's stuck in one place. A system prompt lives inside one app or one codebase. Write a great one for ChatGPT and Claude still knows nothing about you. The same goes for every tool you'll try next year.

It goes stale. A deadline from March gets followed as confidently as one from this week. Someone has to keep it current, and that someone is you.

Where to start

The most useful part of any system prompt is the part about you, and it's also the part no tool can write on your behalf. Your values, your voice, what you're working on, the people who matter. That material doesn't change when you switch tools, so it shouldn't live inside one.

So write it once, as plain text you own, and paste the relevant parts into whichever box each tool gives you. RUMO builds that material as six short documents called context anchors.

If you only write one, start with the free Personal Constitution. It covers how you think and decide, which is the context every system prompt is missing.

The model will still read its instructions before your question. You just get to decide what's in them.

Frequently Asked Questions

What is a system prompt?
A system prompt is a set of standing instructions an AI model reads before it reads your message. It tells the model who it is, who it's serving, what rules to follow, and how to format answers. You usually don't see it, but it shapes every response in the conversation, which is why the same model can behave very differently inside two different apps.
What is the difference between a system prompt and a user prompt?
A system prompt sets the standing rules for the whole conversation. A user prompt is the message you type for the task in front of you right now. The system prompt is written once and applies every turn. The user prompt changes every turn. When the two conflict, AI providers generally design their models to give the standing instructions more weight.
Are custom instructions the same as a system prompt?
They do the same job for regular users. Custom instructions, Claude Projects, and Gemini Gems let you write standing instructions that apply to every chat, the way a developer writes a system prompt. The difference is control. The app still runs its own instructions alongside yours, and you only edit your part. Developers using the API write the whole thing.
Can I see ChatGPT's or Claude's system prompt?
Anthropic publishes the system prompts it uses in the Claude web and mobile apps, with dated versions for each model, and notes that these do not apply to the Claude API. Many apps don't publish theirs. What you can always see and edit is your own layer: custom instructions, project instructions, or a file like CLAUDE.md.
What should I put in a system prompt?
Start with a role, then the context the model can't guess: who it's working for, what that person does, and what's current. Add a short list of rules you could check, the format you want back, and one example of a good answer. Keep it short and plain. A clear page beats a long, clever one, and it's easier to keep current.

Chad Stamm

Chad Stamm

Founder of RUMO

Chad is an AI strategist and integrator, context engineer, and creative director. He built RUMO so your AI can finally work on your behalf, not just answer your questions.

Start free

Give your AI a place to start.

Your Personal Constitution is the first context anchor, and it's free. Build it once, drop it into any AI tool, and watch it stop guessing who you are.

Build your free ConstitutionNo credit card required

Stay on course

Get AI tips, in your inbox.

We'll send practical context tips to help you build the best AI agents and systems.

No spam. Unsubscribe anytime.