What Is a System Prompt and How Do You Write One?

A system prompt is a set of instructions given to an AI model before any conversation begins — it defines the model's role, tone, constraints, and context for everything that follows. Unlike the messages you type in a chat, the system prompt runs silently in the background and shapes every response the model gives.

What this post covers:

• A system prompt is the AI's standing brief — it determines how the model behaves across an entire session, not just for one message.
• System prompts work differently from user prompts: the user prompt asks a specific question, the system prompt sets the rules for how every question gets answered.
• The structure of an effective system prompt follows four elements: role, task, context, and constraints.
• Teams that store system prompts in a shared library — rather than rewriting them per session — get consistent outputs regardless of who runs the prompt.

What Is the Difference Between a System Prompt and a User Prompt?

A user prompt is the message you type each time you want the AI to do something. A system prompt is the instruction layer that sits above it — set once, active for the entire session. Think of the system prompt as the job description and the user prompt as the specific task assigned that day.

When you open ChatGPT and type "summarize this document," that is a user prompt. The system prompt — either set by OpenAI, by a custom GPT's configuration, or by you via the API — determines how ChatGPT approaches that task: how formal its tone is, whether it adds caveats, how long its answers run, and what it refuses to do.

The hierarchy matters: system prompts set default behavior, but final output results from resolving system instructions, user input, and safety constraints together.

‍

Where Do System Prompts Live in Each Major AI Platform?

System prompts behave slightly differently depending on which platform you are using.

In ChatGPT, you access system-level instructions through two routes. For everyday use, go to Settings → Personalization → Custom Instructions — two fields let you tell ChatGPT about your role and how you want it to respond. These apply to every conversation until you change them. For custom GPTs, the "Instructions" field is the full system prompt. Via the API, you pass a system prompt in the system parameter of your request.

In Claude, system prompts are a first-class feature — Anthropic's own documentation recommends using them for role assignment, output formatting, and guardrails. Via the API, you pass your system prompt in the system parameter. Claude is particularly good at following detailed system prompts precisely, making it well-suited for workflows where consistency matters.

In Gemini, system instructions are set in the API via the system_instruction parameter. In Google AI Studio, you can set a system prompt in the interface before testing. Gemini Advanced users can add persistent instructions through the settings panel. Gemini's system prompts tend to work best when they are concise and direct — shorter than what Claude or GPT typically need.

‍

What Makes a Good System Prompt?

Effective system prompts share four elements, regardless of which model you use:

Role. Tell the model who it is. "You are a customer support specialist for a SaaS company" is more useful than "be helpful" because it gives the model a consistent reference point for every response.

Task. Describe what the model's primary job is in this context. "Your job is to answer product questions based on the documentation provided, and escalate to a human agent when a query falls outside that scope."

Context. Provide the information the model needs to do the job well — your company name, your audience, your product, your tone. This is where connecting your organization's own data directly to the prompt produces significantly better outputs than relying on generic AI responses.

Constraints. Define what the model should not do. "Do not speculate about pricing. Do not provide legal advice. Always respond in under 150 words." Constraints prevent the model from drifting outside the intended use case, which matters most in team workflows where multiple people are using the same prompt.

‍

How Do System Prompts Behave Differently Across GPT, Claude, and Gemini?

System prompts work with GPT-4o, GPT-4, Claude, and Gemini, and most concepts transfer across all major models. But each model has tendencies worth knowing.

Claude responds best to explicit XML structure, direct instructions without excessive politeness framing, and detailed system prompts. GPT handles more ambiguous prompts but benefits from clear output formatting instructions and step-by-step task framing when needed. Gemini models respond well to multimodal contexts and benefit from explicit grounding instructions when working with retrieved documents.

In practice: write your system prompt once for the model you use most, then test it across others. Most well-structured system prompts transfer with minor adjustments rather than requiring a complete rewrite.

‍

How Do Teams Store and Reuse System Prompts?

The problem most teams hit is not writing a good system prompt — it is finding it again the next time they need it, and ensuring teammates use the same tested version rather than writing their own variation from scratch.

A system prompt stored only in a chat session disappears with that session. One saved in a personal document never reaches the colleague who needs it. Two people independently writing system prompts for the same task almost always produce prompts that give different outputs — which means the team's AI outputs are inconsistent by default.

A shared prompt library solves this. In Promptitude, system prompts live in the same shared library as all other prompts, tagged by department or workflow so anyone on the team can find and use the tested version. When a system prompt is improved, every team member gets the updated version immediately — rather than each person maintaining their own copy.

Store, tag, and share your team's system prompts in one place. Build your shared prompt library in Promptitude →‍

Frequently Asked Questions

Can I use the same system prompt in ChatGPT, Claude, and Gemini?

Mostly yes, with minor adjustments. The core instructions — role, task, context, constraints — transfer across models. Claude tends to follow detailed instructions more precisely and responds well to XML structure. Gemini works best with shorter, more direct prompts. GPT handles more ambiguous wording but benefits from explicit output format specifications. Store one base version and note any model-specific adjustments as variants.

How long should a system prompt be?

Most effective system prompts are 50–300 words. Keep it as short as possible while still being complete. Shorter is better for simple persona or tone shifts. Longer is necessary when you need specific formatting rules, constraints, or multi-step behaviors. Avoid making them so long that the model ignores parts of it — a 200-word system prompt uses roughly 250–300 tokens from your context window.

Does a system prompt affect every message in a conversation?

Yes. The system prompt is active for the entire session. Every user message is interpreted through the lens of the system prompt's instructions. If you want different behavior for a specific message, you either change the system prompt or override specific instructions within the user prompt itself.

What is the difference between a system prompt and custom instructions in ChatGPT?

Custom Instructions in ChatGPT are a consumer-friendly interface for setting persistent preferences — they get injected into the system prompt automatically. When you use the OpenAI API directly, you write the system prompt yourself and have full control over its content. Custom Instructions are the simplified version; the API system prompt is the full version.

‍