What is a system prompt?

A system prompt is the initial instruction given to an AI model that defines its behavior, personality, and constraints. It's set before any user messages and shapes how the model responds throughout the conversation.

How do I write a good system prompt?

Start with a clear role definition, add specific constraints and rules, define the expected output format, and include examples if needed. Keep it concise but thorough — every word in a system prompt affects the model's behavior.

Do all AI models support system prompts?

Most modern LLMs support system prompts, including ChatGPT (GPT-4, GPT-5), Claude, Gemini, and Llama. The exact format varies by API, but the concept is universal.

How long should a system prompt be?

It depends on your use case. Simple chatbots might need 50-100 words. Complex agents with specific behaviors can use 500-2000 words. Longer prompts use more tokens per request, so balance thoroughness with cost.

Can I save my system prompts?

Your last prompt is automatically saved in your browser's local storage. Use the Copy button to save prompts to your clipboard for permanent storage elsewhere.

System Prompt Builder

What Are System Prompts?

If you’ve used ChatGPT, Claude, or any other LLM through an API, you’ve probably encountered system prompts — even if you didn’t realize it. A system prompt is the hidden instruction that runs before your conversation starts. It’s what turns a general-purpose language model into a specific assistant, persona, or tool.

When you chat with an AI through a consumer app, there’s almost always a system prompt working behind the scenes. It might say something like “You are a helpful assistant” or it could be thousands of words defining exact behaviors, output formats, and guardrails.

Why System Prompts Matter

The system prompt is the single biggest lever you have over an LLM’s behavior. A well-crafted system prompt can:

Define expertise: Turn a generalist model into a domain specialist (legal advisor, code reviewer, medical triage assistant)
Set boundaries: Prevent the model from answering questions outside its scope
Control format: Ensure responses come back as JSON, markdown tables, bullet points, or any structure you need
Establish tone: Make the model formal, casual, terse, or verbose
Reduce hallucination: By giving the model explicit instructions about when to say “I don’t know”

Skip the system prompt and you’re leaving the model’s behavior up to chance. That might be fine for casual use, but it’s a non-starter for production applications.

Anatomy of an Effective System Prompt

Most good system prompts share a common structure, which is exactly what this builder helps you assemble:

Role Definition

Start with who the AI is. “You are a senior Python developer who specializes in Django applications.” This anchors the model’s responses in a specific expertise area.

Task Description

What should the model actually do? “Your job is to review code snippets and identify bugs, security vulnerabilities, and performance issues.” Be specific — vague tasks produce vague results.

Constraints and Rules

This is where you set guardrails. “Never suggest deprecated APIs. Always include error handling in code examples. If you’re unsure about something, say so instead of guessing.” Constraints prevent the model from going off the rails.

Output Format

Tell the model exactly how to structure its responses. “Respond with a markdown list where each item includes the issue severity (High/Medium/Low), a description, and a suggested fix.” Without format instructions, you’ll get inconsistent output.

Tone and Persona

“Be direct and technical. Skip pleasantries. Use code comments instead of prose explanations when possible.” Tone instructions shape the feel of every response.

Examples

Few-shot examples inside system prompts are incredibly powerful. Show the model one or two input/output pairs that demonstrate exactly what you want, and it’ll pattern-match from there.

Common System Prompt Patterns

The Specialist: Define deep expertise in a narrow domain. Works well for customer support bots, technical advisors, and tutoring applications.

The Formatter: Minimal personality, heavy format instructions. Great for data extraction, API response generation, and structured output tasks.

The Guardrailed Assistant: Lots of “do not” rules and boundary conditions. Essential for customer-facing applications where you can’t afford unexpected responses.

The Chain-of-Thought Reasoner: Instruct the model to think step-by-step before answering. Improves accuracy on complex reasoning tasks at the cost of longer responses.

Tips for Better System Prompts

Be specific over general. “Respond in 2-3 sentences” beats “be concise.”
Test iteratively. Write a draft, try edge cases, refine. System prompts rarely work perfectly on the first try.
Use positive instructions. “Always cite sources” works better than “don’t make unsourced claims” — models follow positive instructions more reliably.
Keep it organized. Use headers, numbered lists, and clear sections. Models parse structured prompts better than walls of text.
Watch your token count. Every token in the system prompt is sent with every request. A 2,000-token system prompt on GPT-5.4 at $10/M input tokens costs $0.02 per request before the user even types anything.

System Prompts Across Different Models

Each model handles system prompts slightly differently:

OpenAI (GPT-4, GPT-5): Uses a dedicated system role in the messages array. Developer messages can add additional system-level instructions.
Anthropic (Claude): Has a separate system parameter outside the messages array. Supports longer system prompts without eating into the conversation context.
Google (Gemini): Uses systemInstruction in the generation config.
Open-source models (Llama, Mistral): Typically use chat templates with a system token. The exact format varies by model and serving framework.

Despite the API differences, the content of a good system prompt is universal. Write it once with this builder, then adapt the delivery mechanism to your model of choice.