How to Write System Prompts for Claude and ChatGPT (2026 Guide) | AI Prompts Pro

Master system prompts for Claude and ChatGPT. Learn how to set context, define behavior, and create AI assistants that consistently deliver exactly what you need.

2026-02-12

How to Write System Prompts for Claude and ChatGPT

Published February 12, 2026

Max Sterling

February 12, 2026 · 15 min read

System prompts are the secret weapon of power users who get consistently amazing results from AI. While most people write one-off prompts and hope for the best, system prompts let you configure an AI's entire personality, knowledge boundaries, and output style upfront - then use it like a specialized assistant who already knows exactly what you need.

Think of system prompts as the difference between hiring a general contractor versus someone who's worked on 50 projects exactly like yours. The system prompt is where you define the expertise, constraints, and behavior patterns that make an AI assistant truly useful instead of just occasionally helpful. If you're new to prompting, start with our prompt engineering fundamentals guide.

What Are System Prompts?

A system prompt (also called system instructions or system message) is a persistent set of instructions that defines how an AI model should behave across an entire conversation. Unlike user prompts that request specific outputs, system prompts set the baseline for who the AI is, how it should respond, and what rules it should follow.

System prompts run before every user message, but the user doesn't see them. They're the hidden instruction layer that shapes every response the AI generates. When you create a Custom GPT in ChatGPT or use Claude Projects, you're writing system prompts.

System Prompts vs. User Prompts

System Prompt (Persistent):
"You are an expert Python developer who specializes in data science. Always provide working code with comments. Prioritize readability over cleverness. When suggesting libraries, explain why you chose them."

User Prompt (One-time):
"Write a function that calculates the correlation between two lists."

The system prompt runs invisibly before the user prompt, shaping how the AI interprets and responds to the request. Without the system prompt, you'd get a generic function. With it, you get code that matches your style preferences, includes explanations, and uses libraries you care about.

Why System Prompts Matter in 2026

As AI models get more capable, the challenge isn't what they can do - it's getting them to do exactly what you want, consistently, without re-explaining your preferences every time. System prompts solve this by creating specialized AI assistants for your specific workflows.

Here's what changes when you master system prompts:

Consistency: Stop getting different quality outputs depending on how you phrase things
Efficiency: No more re-explaining context or preferences in every conversation
Specialization: Create AI assistants that are experts in your specific domain
Control: Define boundaries, tone, format, and behavior upfront
Quality: Get outputs that match your standards without extensive editing

The Anatomy of a Great System Prompt

A well-structured system prompt has five core components. You don't always need all five, but understanding each helps you build exactly what you need.

1. Role and Identity

Who is this AI? What's their expertise? This sets the knowledge domain and perspective the AI should adopt.

You are a senior technical writer at a B2B SaaS company with 10 years of experience. You specialize in turning complex technical concepts into clear documentation that non-technical users can understand and act on.

The more specific the role, the better the AI can match that expertise level and perspective. "You are a marketer" is weak. "You are a performance marketing specialist focused on e-commerce brands, with deep expertise in Facebook Ads and Google Shopping campaigns" is much stronger.

2. Knowledge Boundaries and Context

What should the AI know about your specific situation? What context shapes how it should respond?

Our company builds project management software for construction teams. Our users are typically superintendents and project managers who are tech-capable but not developers. Our product emphasizes mobile-first workflows and offline functionality. Our brand voice is professional but approachable - like a knowledgeable colleague, not a stuffy consultant.

This context prevents the AI from giving generic advice. It knows your constraints, audience, and brand personality.

3. Behavioral Guidelines

How should the AI approach tasks? What's its decision-making framework?

When answering questions: - Always ask clarifying questions if the request is ambiguous - Provide specific examples rather than abstract explanations - If you're uncertain about something, say so explicitly - Prioritize actionable advice over theoretical knowledge - Use bullet points and headers for scannability - Keep responses under 300 words unless specifically asked for more detail

4. Output Format and Style

What should the output look like? This is where you define formatting, structure, tone, and style guidelines.

Output format requirements: - Use markdown formatting (headers, bullets, code blocks) - Start with a brief summary sentence - Use short paragraphs (2-3 sentences maximum) - Include examples wherever possible - End with a "Next Steps" section if the response is longer than 200 words - Tone: Conversational but professional. Use contractions. Avoid jargon unless explaining it. - Never use corporate buzzwords like "synergy," "circle back," or "move the needle"

5. Constraints and Guardrails

What should the AI NOT do? Constraints are as important as capabilities.

Important constraints: - Never make up statistics or data - if you don't have verified information, say so - Don't recommend tools or services without explaining why they're appropriate for this use case - Avoid overly complex solutions when simpler ones exist - Don't use technical terms without defining them first - If a request goes outside your area of expertise, acknowledge it and provide relevant resources instead

System Prompts for Different Use Cases

Let's look at real-world system prompt examples you can adapt for your needs. For more specialized prompts, check out our business-focused ChatGPT prompts.

Content Editor and Reviewer

You are an experienced content editor who has worked at top publications like The Atlantic and Wired. Your job is to review written content and provide specific, actionable feedback to make it better. When reviewing content: 1. Start with what's working well (2-3 specific strengths) 2. Identify the biggest issue that's holding the piece back 3. Provide 3-5 specific improvements, ranked by impact 4. Point out clarity issues, logical gaps, or unsupported claims 5. Suggest a stronger headline if the current one is weak 6. Note pacing problems (too slow, too rushed, uneven) Your feedback should be direct but constructive. Focus on substance over style. If something is confusing, explain why and how to fix it. Always provide examples of how to improve, not just what's wrong. Format feedback as: ## Strengths [bullet list] ## Main Issue [explanation] ## Improvements [numbered list with specific suggestions] ## Revised Headline Suggestion [if needed]

Code Review Assistant

You are a senior software engineer who specializes in code reviews. Your focus is on writing maintainable, testable code that follows best practices. When reviewing code: - Check for potential bugs or edge cases - Identify performance issues or inefficiencies - Suggest more readable alternatives to complex code - Point out security concerns - Recommend better naming for unclear variables/functions - Note missing error handling - Suggest opportunities for refactoring Your feedback should be: - Specific: Point to exact lines or patterns - Explanatory: Explain WHY something is an issue - Constructive: Provide better alternatives, not just criticism - Prioritized: Mark issues as Critical, Important, or Optional Format output as: ## Critical Issues [issues that could cause bugs or security problems] ## Important Improvements [things that hurt maintainability or performance] ## Optional Enhancements [nice-to-haves that would make the code better] For each issue, provide a code example of the fix when possible.

Strategic Business Advisor

You are a strategic business consultant with 15 years of experience advising startups and mid-sized companies. You specialize in growth strategy, market positioning, and operational efficiency. Your approach: - Ask clarifying questions before jumping to recommendations - Consider both short-term wins and long-term strategy - Balance ambition with practical constraints - Base advice on first principles, not trends - Call out assumptions in your recommendations - Provide frameworks, not just opinions When giving strategic advice: 1. Restate the challenge to ensure understanding 2. Identify 2-3 key factors that will determine success 3. Present 2-3 strategic options with pros/cons 4. Recommend the best path forward with clear reasoning 5. Outline 3-5 specific next steps to take in the next 30 days Your tone is confident but not prescriptive. You offer informed opinions while acknowledging uncertainty. You think in terms of trade-offs, not perfect solutions. Avoid: Generic advice, buzzwords, recommendations without context about implementation difficulty

Email Response Assistant

You are a professional communication specialist who helps write clear, effective email responses. Context: You're writing on behalf of a [YOUR JOB TITLE] at [YOUR COMPANY TYPE]. Response guidelines: - Match the tone of the incoming email (formal if formal, casual if casual) - Keep responses under 150 words unless the situation demands more - Lead with the most important information - Use active voice and clear language - If declining a request, be polite but direct - End with a clear next step or closing Email structure you follow: 1. Brief greeting (if first email in thread) 2. Address their main point/question first 3. Provide necessary details 4. Clear next step or closing 5. Professional sign-off When given an email to respond to: - First, summarize what they're asking for - Then provide the response - If the request is unclear, draft questions to clarify before responding

Advanced System Prompt Techniques

Chain-of-Thought Reasoning

For complex tasks, instruct the AI to think step-by-step before responding. This dramatically improves reasoning quality.

Before providing your final response, think through the problem step-by-step: 1. What is the user really asking for? (Restate in your own words) 2. What information do you need to answer well? 3. What are 2-3 possible approaches? 4. Which approach is best and why? 5. What could go wrong with this recommendation? Show this thinking process in a "Reasoning" section, then provide your final recommendation separately.

Self-Critique and Improvement

Tell the AI to evaluate its own outputs before finalizing them.

After drafting a response: 1. Check if it directly answers the user's question 2. Verify all claims are accurate or clearly marked as uncertain 3. Ensure the response is actionable, not just informational 4. Confirm the tone matches the guidelines 5. If any of these checks fail, revise before providing the final response

Dynamic Formatting Based on Complexity

Instruct the AI to adapt its format based on the complexity of the request.

Adjust your response format based on complexity: Simple questions (1-2 sentence answers): - Provide a direct answer - Optional: One brief example Medium complexity (requires explanation): - Brief intro sentence - 3-5 bullet points with details - Example if helpful Complex questions (multi-faceted issues): - Executive summary (1-2 sentences) - Sections with clear headers - Examples for each major point - "Next Steps" section at the end

System Prompts: Claude vs ChatGPT

While the core principles work across both platforms, there are subtle differences in how Claude and ChatGPT interpret system prompts. Want a deeper comparison? Read our Claude vs ChatGPT prompting guide.

ChatGPT System Prompt Characteristics

Works well with structured, explicit instructions
Responds strongly to personality and role definitions
Good at following format specifications precisely
Can be verbose - use length constraints if needed
Strong with creative and generative tasks

Claude System Prompt Characteristics

Excels at nuanced, contextual understanding
More cautious by default - benefits from explicit permission to be direct
Handles long, complex system prompts very well
Strong with analytical and reasoning tasks
Better at "reading between the lines" of conversational requests

Example: Same Goal, Different Optimization

For ChatGPT (more explicit structure):

You are a marketing copywriter. For every request, follow this exact process: 1. Identify the target audience 2. Determine the primary benefit 3. Choose a persuasive framework (PAS, AIDA, etc.) 4. Write the copy 5. Provide 2 alternative variations Format all responses with clear section headers.

For Claude (more contextual):

You're a marketing copywriter who understands persuasion psychology. When writing copy, consider what the audience cares about most, lead with the strongest benefit, and use proven frameworks like problem-agitate-solution when appropriate. Always provide alternatives so there are options to choose from. Be clear about which approach you think will work best and why.

Common System Prompt Mistakes

Mistake 1: Too Vague

❌ Bad: "You are helpful and knowledgeable."

✅ Good: "You are a senior data analyst with expertise in e-commerce analytics. You help marketing teams understand their customer data and make data-driven decisions about campaigns and product positioning."

Mistake 2: Conflicting Instructions

Make sure your instructions don't contradict each other:

❌ Bad: "Be extremely detailed and thorough. Keep all responses under 100 words."

✅ Good: "Provide detailed explanations when the topic is complex, but keep responses under 200 words unless the user specifically requests more depth."

Mistake 3: Overloading with Rules

More instructions isn't always better. A 2000-word system prompt with 50 rules will confuse the AI. Focus on the 5-10 most important guidelines.

Mistake 4: Not Testing and Iterating

Your first system prompt won't be perfect. Test it with various requests, see where it falls short, and refine. System prompt development is iterative.

Mistake 5: Forgetting About Token Limits

System prompts count against your token limit. If you're working with long documents or conversations, keep your system prompt focused and efficient.

Ready-Made System Prompts for Every Use Case

Why start from scratch? Get instant access to 200+ tested system prompts for Custom GPTs and Claude Projects.

Get System Prompts

Building Your System Prompt Library

Don't write system prompts from scratch every time. Build a personal library of proven templates you can adapt.

Start with These Core Templates

Expert Consultant: For strategic advice in your domain
Editor/Reviewer: For improving written content
Technical Assistant: For code, documentation, or technical writing
Research Analyst: For gathering and synthesizing information
Creative Partner: For brainstorming and ideation

Create one solid template for each category, then customize the specifics based on the project.

Version Control Your System Prompts

As you refine a system prompt, keep track of what worked and what didn't. I keep a simple doc with:

Current version of the system prompt
Changelog (what I changed and why)
Example conversations showing it working well
Known limitations or edge cases

System Prompts for Custom GPTs and Claude Projects

Both ChatGPT's Custom GPTs and Claude Projects are powered by system prompts. Here's how to use them effectively.

Custom GPTs (ChatGPT)

When creating a Custom GPT, you're essentially writing a system prompt in the "Instructions" field. Best practices:

Use the "Knowledge" feature to upload reference documents
Enable web browsing if the GPT needs current information
Test with diverse prompts to ensure consistent behavior
Use "Conversation starters" to guide users on how to use your GPT effectively

Claude Projects

Claude Projects let you set custom instructions and upload reference materials. They're perfect for ongoing work where context matters:

Use the custom instructions to define the assistant's role and approach
Upload style guides, documentation, or reference materials
Claude handles long context exceptionally well - use it
Projects maintain conversation history, so you can build on previous work

Testing and Refining Your System Prompts

A system prompt isn't finished when you write it - it's finished when it consistently produces the outputs you need.

The Testing Process

Write Version 1: Start with the core components (role, guidelines, format)
Test with 5-10 realistic prompts: Use actual requests you'd make, not toy examples
Identify patterns in failures: Where does it consistently miss the mark?
Refine the weak areas: Add specific instructions addressing the failure modes
Retest: Did your changes fix the issues without breaking what worked?
Repeat: Keep iterating until you're getting consistent, high-quality outputs

Questions to Ask During Testing

Does the AI understand its role and expertise boundaries?
Are outputs formatted consistently the way I want?
Is the tone appropriate for the use case?
Does it ask clarifying questions when needed, or guess?
Are responses too long/short/detailed/vague?
Does it follow constraints I set?
Would I need to edit this significantly before using it?

The Future of System Prompts

As AI models evolve, system prompts will become even more important. We're moving toward a world where you'll have dozens of specialized AI assistants - each configured for a specific role through carefully crafted system prompts.

The skill of writing effective system prompts will be as valuable as writing good code or compelling copy. It's how you'll configure AI to work exactly the way you need it to, without constant micromanagement.

Start Building Your AI Assistant

You now understand the anatomy of powerful system prompts. The next step is applying this to your actual work. Here's what I recommend:

Identify your most frequent AI use case: What do you ask AI to do repeatedly?
Write a system prompt for that use case: Use the five-component framework
Test it with 10 real requests: See where it works and where it breaks
Refine based on what you learn: Fix the weak spots
Save and reuse it: Build your prompt library

System prompts aren't magic - they're just clear communication about what you need from an AI assistant. The better you get at writing them, the more valuable AI becomes as a tool in your workflow. For more advanced techniques, explore our best ChatGPT prompts collection.

Want access to a library of professionally crafted system prompts you can use immediately? AI Prompts Pro includes 200+ system prompt templates for Custom GPTs, Claude Projects, and API integrations. Start building your AI assistant army today. Need help? Check our FAQ page for answers.