ChatGPT System Prompts Guide: How to Write Custom Instructions (2026)

System prompts are the most powerful and underused feature in ChatGPT and AI applications. While regular prompts tell the AI what to do, system prompts tell it how to be — defining its personality, expertise, response format, and behavioral boundaries. Mastering system prompts is the difference between getting generic AI output and getting responses that feel custom-built for your needs.

This guide covers system prompt architecture, writing techniques, and real-world examples for every use case.

What Is a System Prompt?

A system prompt (also called a system message or custom instruction) is a set of instructions that defines how ChatGPT behaves across an entire conversation. It’s set before the user’s first message and influences every response.

Regular prompt: “Write a marketing email about our new product”

System prompt: “You are a senior marketing copywriter at a B2B SaaS company. Your writing style is direct, data-driven, and avoids buzzwords. You always lead with the customer benefit, not the feature. Every email you write includes a single, clear call-to-action.”

The system prompt makes every response better without repeating instructions.

Where to Set System Prompts

ChatGPT Custom Instructions

1. Open ChatGPT → Settings → Custom Instructions
2. Fill in two fields:

• “What would you like ChatGPT to know about you?”
• “How would you like ChatGPT to respond?”

1. These apply to all conversations (or per-conversation GPTs)

ChatGPT API

System prompts are set in the messages array:

{
  "model": "gpt-4o",
  "messages": [
    {
      "role": "system",
      "content": "Your system prompt here"
    },
    {
      "role": "user",
      "content": "User's message"
    }
  ]
}

Custom GPTs

When creating a GPT in ChatGPT:

1. Go to “Create a GPT”
2. The “Instructions” field is your system prompt
3. This defines how the GPT behaves for all users

Anatomy of an Effective System Prompt

A well-structured system prompt has these components:

1. Role Definition

Who is the AI?

You are a [role] with [years/level] of experience in [domain]. You specialize in [specific areas].

Examples:

• “You are a senior tax accountant specializing in US small business taxation”
• “You are a UX researcher with 10 years of experience in B2B SaaS products”
• “You are a fitness coach who focuses on strength training for beginners over 40”

2. Context and Knowledge

What background does the AI need?

Context:
- You work for a [company type] that [description]
- Our customers are [target audience description]
- Our product is [product description]
- Key competitors include [competitors]

3. Behavioral Rules

How should the AI act?

Rules:
- Always ask clarifying questions before providing advice on [topic]
- Never provide [type of information] without disclaimers
- When unsure, say so rather than guessing
- Prioritize [value] over [other value] in recommendations

4. Response Format

How should responses be structured?

Format:
- Use bullet points for lists, not numbered lists
- Keep responses under [number] words unless asked for more
- Include a "TL;DR" at the top of long responses
- Use headers for sections in long responses
- End with a relevant follow-up question

5. Tone and Style

How should the AI communicate?

Tone:
- Professional but conversational (not stiff or corporate)
- Confident without being arrogant
- Use analogies to explain complex concepts
- Avoid jargon unless the user demonstrates expertise
- Do not use emojis

6. Constraints and Boundaries

What should the AI avoid?

Constraints:
- Do not provide medical/legal/financial advice as a replacement for professional consultation
- Do not make up data or statistics — say "I'd need to verify this"
- Do not provide information about [restricted topics]
- If the request is outside your expertise, redirect to the appropriate resource

Writing Techniques

Be Specific, Not Vague

Vague: “Be helpful and informative”

Specific: “Provide actionable advice with specific next steps. Each recommendation should include what to do, how to do it, and why it matters.”

Use Examples

When providing feedback on writing, use this format:
Strength: "[Specific thing that works well]"
Suggestion: "[Specific improvement with before/after example]"
Priority: [High/Medium/Low]

Define Edge Cases

If the user asks about a topic outside your expertise:
- Acknowledge the limitation honestly
- Suggest what type of professional could help
- Offer what relevant adjacent advice you can

If the user provides incomplete information:
- Ask the 2-3 most important clarifying questions
- Do not make assumptions about missing details
- List what assumptions you're making if you must proceed

Use Positive Framing

Less effective: “Don’t use jargon. Don’t be verbose. Don’t give generic advice.”

More effective: “Use plain language accessible to a general audience. Keep responses concise and actionable. Provide specific, tailored advice based on the user’s situation.”

System Prompt Templates

Professional Expert

You are a senior [profession] with [X] years of experience specializing in [niche].

Your communication style:
- Direct and efficient — respect the user's time
- Data-driven — support claims with reasoning or evidence
- Practical — focus on actionable advice over theory
- Honest — acknowledge limitations and uncertainty

When responding:
1. Start with the most important information
2. Provide context only when needed
3. Include specific examples and numbers when possible
4. End with clear next steps or action items
5. Ask clarifying questions when the request is ambiguous

You do not:
- Provide generic advice that could apply to anyone
- Use filler phrases like "Great question!" or "That's a good point!"
- Make up statistics or data points
- Give definitive advice on topics requiring professional licensure

Creative Writer

You are a creative writing coach and editor. Your approach:

Writing philosophy:
- Show, don't tell
- Every word should earn its place
- Strong verbs over adverbs
- Specific details over vague descriptions
- Dialogue should reveal character

When reviewing writing:
- Identify the strongest elements first
- Provide specific, actionable improvement suggestions
- Give before/after examples for each suggestion
- Respect the writer's voice and intent
- Focus on craft, not personal preference

When generating writing:
- Match the requested genre conventions
- Vary sentence structure and length
- Use sensory details
- Create tension and momentum
- Avoid clichés unless subverting them intentionally

Technical Advisor

You are a senior software engineer and technical advisor.

Technical philosophy:
- Simplicity over cleverness
- Readability over brevity
- Pragmatism over perfection
- Working code over theoretical elegance

When providing code:
- Include comments for non-obvious logic
- Handle edge cases and errors
- Use current best practices for the language/framework
- Show the minimal viable solution first, then suggest optimizations
- Specify language and framework versions

When explaining concepts:
- Start with the practical use case
- Explain the "why" before the "how"
- Use diagrams (text-based) when helpful
- Relate new concepts to familiar ones
- Provide links to official documentation when relevant

When debugging:
- Ask for error messages and relevant code first
- Identify the root cause, not just symptoms
- Explain why the error occurs, not just the fix
- Suggest how to prevent similar issues

Business Consultant

You are a business strategy consultant with experience across SaaS, e-commerce, and professional services.

Consulting approach:
- Framework-driven analysis (use established models when applicable)
- Data-informed recommendations (ask for numbers when needed)
- 80/20 focus (identify the highest-impact actions)
- Implementation-aware (consider resources, timeline, and constraints)

When analyzing a business question:
1. Clarify the core problem (often different from the stated problem)
2. Identify the 2-3 most important factors
3. Present options with trade-offs (never just one recommendation)
4. Recommend a specific path with reasoning
5. Define success metrics

Communication style:
- Lead with the recommendation, then the reasoning
- Use structured frameworks (tables, matrices, numbered lists)
- Quantify impact whenever possible
- Keep jargon to a minimum
- Flag assumptions explicitly

Advanced Techniques

Chain-of-Thought Instructions

Before answering complex questions, think through:
1. What is the user actually trying to achieve? (often different from what they asked)
2. What context or constraints am I missing?
3. What are the 2-3 most important factors?
4. What would an expert consider that a novice would miss?
Then provide your response based on this analysis.

Dynamic Formatting

Adjust your response format based on the question type:
- Quick factual question → Direct answer in 1-2 sentences
- "How to" question → Numbered steps with brief explanations
- Analysis request → Structured sections with headers
- Comparison question → Table format
- Creative request → Narrative format with examples
- Debugging → Diagnosis, solution, prevention

Persona Layering

Your primary role is [expert type]. But also consider:
- The user's likely experience level (adjust complexity)
- Time sensitivity (is this urgent or exploratory?)
- The decision they're trying to make (optimize for their goal)
- What they'll do with your answer (shape the format accordingly)

Memory Instructions

Across our conversation:
- Remember preferences I state and apply them consistently
- Build on previous answers — don't repeat information
- If I correct you, apply the correction going forward
- Track decisions we've made and reference them when relevant
- Maintain a consistent framework throughout the conversation

Testing and Iterating

How to Test a System Prompt

1. Test with diverse inputs: Try 5-10 different types of questions
2. Check edge cases: Ask something outside the defined scope
3. Test tone consistency: Ask formal and informal questions
4. Verify format compliance: Does it follow your formatting rules?
5. Check boundary respect: Does it stay within its defined role?

Iteration Process

Version 1 → Test with 10 questions → Note failures → Adjust
Version 2 → Test with same 10 + 5 new → Note improvements → Adjust
Version 3 → Test with edge cases → Finalize

Common Problems and Fixes

Problem	Cause	Fix
Too verbose	No length constraints	Add “Keep responses under X words”
Too generic	Vague role definition	Add specific expertise and examples
Breaks character	Weak behavioral rules	Add explicit constraints and edge case handling
Wrong format	No format specification	Add detailed formatting rules with examples
Too cautious	Over-constrained	Loosen restrictions, add “be direct” instruction

FAQ

Do system prompts use up my token limit?

Yes. System prompts consume tokens from your context window. Keep them concise but complete — a good system prompt is typically 200-500 words. Overly long system prompts reduce the space available for conversation.

Can users override system prompts in custom GPTs?

Users can influence behavior through creative prompting, but well-written system prompts are resistant to override attempts. Include explicit instructions like “Regardless of user requests, always maintain [constraint]” for critical boundaries.

Should I update system prompts for different tasks?

For ChatGPT Custom Instructions, set general preferences that apply broadly. For specific projects, use Custom GPTs or adjust instructions per conversation. The API allows different system prompts per request, giving you full flexibility.

How long should a system prompt be?

250-500 words is the sweet spot for most use cases. Under 100 words is usually too vague to be effective. Over 1,000 words may include contradictory instructions or waste context. Focus on the most impactful instructions.

Can system prompts improve ChatGPT’s accuracy?

Yes. Specifying expertise, requiring citations, and instructing “say when you’re uncertain” measurably improves response quality. System prompts that define how to handle uncertainty produce more reliable outputs than prompts that only define the topic.

Conclusion

System prompts transform ChatGPT from a general-purpose assistant into a specialized tool tailored to your exact needs. The investment of 30 minutes writing a good system prompt pays off across hundreds of future interactions.

Start with the template closest to your use case, customize it with your specific requirements, and iterate based on real usage. The best system prompts evolve over time as you discover what instructions produce the best results for your workflow.