System Prompts Explained: What They Are and How to Use Them in ChatGPT, Claude, and Gemini

Every AI tool you use has a secret layer of instructions you never see.

Before you type a single word into ChatGPT, Claude, or Gemini, a hidden set of rules has already told the AI how to behave, what tone to use, what to refuse, and how to interpret your requests. These hidden instructions are called system prompts, and understanding them changes how you use AI entirely.

Most people treat AI like a search bar. Type a question, get an answer. But the people getting consistently better results? They have figured out that the real power is in controlling the instructions that run before the conversation starts.

This guide breaks down what system prompts are, how they work across every major AI platform, and how to write your own to get dramatically better output from every conversation.

What Is a System Prompt?

A system prompt is a set of instructions given to an AI model before it receives any input from you. Think of it as the AI's job description. It tells the model:

• Who to be — "You are a senior Python developer" or "You are a friendly writing coach"
• How to behave — "Always explain your reasoning" or "Keep responses under 200 words"
• What to prioritize — "Focus on practical examples" or "Cite sources for every claim"
• What to avoid — "Never use jargon" or "Do not make assumptions about the user's skill level"

The system prompt sits above your conversation. Every message you send gets filtered through these instructions first. That is why two people can type the exact same question into the same AI model and get completely different answers — different system prompts, different behavior.

A Simple Example

Without a system prompt, asking "How do I improve my landing page?" gets you a generic marketing textbook answer.

With this system prompt:

You are a conversion rate optimization specialist with 10 years of experience. You have personally run A/B tests on over 500 landing pages. You give specific, actionable advice based on data patterns you have observed. You avoid generic best practices and instead focus on counterintuitive findings. When you make a recommendation, you explain the evidence behind it.

Now the same question produces a response with real depth — specific patterns, data-backed recommendations, and the kind of nuance you would get from an expensive consultant.

The difference is not the model. It is the instructions.

System Prompts vs. User Prompts vs. Custom Instructions

These three terms get confused constantly. Here is the difference:

Term	What It Is	Who Sets It	When It Runs
System Prompt	Hidden instructions defining the AI's role and behavior	Developers, or you via API/Projects	Before every conversation
User Prompt	Your actual message or question	You	Each message you send
Custom Instructions	A user-facing version of system prompts	You, via settings	Automatically, every conversation

System prompts are the developer-level concept. When OpenAI builds ChatGPT, they write a system prompt that tells the model how to behave as "ChatGPT." When Anthropic builds Claude, they write one that defines Claude's personality and guidelines.

Custom Instructions are the consumer-friendly version. ChatGPT, Claude, and Gemini all give you a settings panel where you can write persistent instructions that apply to every conversation. Under the hood, these get injected as part of the system prompt.

User prompts are simply what you type in the chat box each time.

The hierarchy matters: system prompt shapes everything, custom instructions personalize within that frame, and your message is interpreted through both layers.

How System Prompts Work on Each Platform

ChatGPT (OpenAI)

ChatGPT gives you two ways to set system-level instructions:

Custom Instructions (Settings → Personalization → Custom Instructions)

Two fields:
1. "What would you like ChatGPT to know about you?" — Context about your role, expertise, preferences
2. "How would you like ChatGPT to respond?" — Output rules, format preferences, tone

These apply to every new conversation until you change them.

GPTs (Custom GPTs)

If you build a custom GPT, you write a full system prompt in the "Instructions" field. This is the most powerful option — you define the entire personality, knowledge base, and behavior. Anyone who uses your GPT gets that experience.

API Access

Developers set system prompts via the system role in API calls. This is the most flexible option but requires technical knowledge.

Claude (Anthropic)

Claude handles system prompts through Projects:

Projects

Each Claude Project has its own system prompt, called "Custom Instructions." This is powerful because you can create separate projects for different tasks — one for code review, one for writing, one for data analysis — each with tailored instructions.

Unlike ChatGPT's single global custom instruction, Claude lets you maintain multiple system prompts and switch between them by switching projects.

API Access

Through the API, you set system prompts in the system parameter. Claude's documentation specifically recommends using system prompts for role assignment, output formatting, and guardrails.

Gemini (Google)

Gemini offers system instructions through:

Gems

Google's version of custom GPTs. Each Gem has its own instruction set that defines its behavior.

API Access

Via the Gemini API, you set a system_instruction that persists across the conversation.

The Cross-Platform Problem

Here is the frustration: you write a great system prompt for Claude, then want to use it in ChatGPT. You have to manually copy it over, adjust it for the platform, and maintain two versions.

This is exactly why keeping your system prompts in an external tool matters. When your prompts live in a dedicated library instead of scattered across platform settings, you can grab any system prompt and paste it into whatever AI tool you are using that day.

How to Write an Effective System Prompt

A good system prompt has five components. You do not need all five every time, but the best ones include most of them.

1. Role Assignment

Tell the AI who it is. Be specific.

Weak: "You are a helpful assistant."

Strong: "You are a technical writer with 8 years of experience documenting APIs for developer audiences. You write in plain English, avoid marketing language, and always include code examples."

The more specific the role, the more the AI draws on relevant knowledge patterns. "Senior Python developer" produces different output than "developer" because the model adjusts its assumptions about what you already know.

2. Task Definition

What should the AI do with every message? Define the core job.

Your job is to review code snippets I share and identify potential bugs, performance issues, and security vulnerabilities. For each issue you find, explain why it is a problem and suggest a specific fix.

3. Output Rules

How should responses look? Set the format.

Format every response as:
- Summary (2-3 sentences)
- Details (bulleted list)
- Action items (numbered list)

Keep responses under 500 words unless I ask for more detail.

4. Behavioral Guardrails

What should the AI avoid? Set boundaries.

Never make up statistics or cite sources you are not confident about. If you are unsure about something, say so explicitly. Do not use corporate jargon, buzzwords, or filler phrases. If my question is ambiguous, ask a clarifying question before answering.

5. Context and Constraints

What background information shapes every interaction?

I work at a B2B SaaS startup with 12 employees. Our tech stack is Python/Django on the backend, React on the frontend. Our audience is mid-market companies in the healthcare space. When you give advice, factor in that we have limited engineering resources and need solutions that are practical for a small team.

Putting It All Together

Here is a complete system prompt for a content strategist:

You are a content strategist who specializes in B2B SaaS companies. You have 10 years of experience growing organic traffic from zero to 100K+ monthly visitors.

When I share a content idea or ask for advice, respond with:
1. Your honest assessment (is this worth pursuing?)
2. The target keyword and estimated difficulty
3. A recommended angle that differentiates from existing content
4. An outline with H2 headers

Be direct. If an idea is bad, say so and explain why. Do not hedge or give wishy-washy feedback. Use specific examples from real companies when possible (you can anonymize them). Assume I understand SEO basics — do not explain concepts like "search intent" or "keyword difficulty" unless I ask.

Notice how every sentence serves a purpose. No filler. No "be helpful and friendly." Just clear instructions that shape behavior.

10 System Prompt Templates You Can Use Today

These are ready to paste into your AI tool of choice. Customize the bracketed sections for your situation.

1. Code Reviewer

You are a senior software engineer conducting a code review. Review code I share for bugs, security issues, performance problems, and readability. For each issue, explain the risk and provide corrected code. Prioritize issues by severity. If the code is clean, say so — do not invent problems.

2. Writing Editor

You are a senior editor at a respected publication. When I share writing, provide specific, actionable feedback on: clarity, structure, argument strength, and tone. Mark exactly which sentences need work and explain why. Do not rewrite my work — point out the problems and let me fix them.

3. Meeting Summarizer

You are an executive assistant skilled at extracting signal from noise. When I share meeting notes or transcripts, produce: (1) Key decisions made, (2) Action items with owners, (3) Open questions that need follow-up, (4) Anything that was discussed but not resolved. Be concise. If something is unclear from the notes, flag it.

4. Email Drafter

You are a professional communication specialist. Write emails that are clear, concise, and appropriate for the context I describe. Default to a professional but warm tone. Never use phrases like "I hope this email finds you well" or "Please do not hesitate to reach out." Keep emails under 150 words unless the situation requires more.

5. Data Analyst

You are a data analyst who explains findings to non-technical stakeholders. When I share data or describe an analysis need, focus on: what the data shows, why it matters, and what action to take. Use plain language. Include the relevant numbers but contextualize them. If the data is insufficient to draw conclusions, say so.

6. Product Manager

You are a senior product manager at a SaaS company. Help me think through product decisions with frameworks, not opinions. When I describe a feature or problem, ask clarifying questions first. Then help me evaluate trade-offs using relevant frameworks (RICE scoring, Jobs to Be Done, opportunity sizing). Challenge my assumptions.

7. Teacher / Explainer

You are a patient, skilled teacher. When I ask about a topic, explain it as if I am smart but completely new to this subject. Start with the core concept, then build up complexity. Use analogies from everyday life. Check my understanding by asking me a question at the end. Never assume prior knowledge.

8. Sales Copywriter

You are a direct-response copywriter who has written for companies like Basecamp and Apple. Write copy that is specific, benefit-focused, and free of buzzwords. Every sentence should either create desire or remove a objection. Do not use superlatives (best, amazing, revolutionary) — use specific proof instead.

9. Research Assistant

You are a research assistant helping me explore a topic thoroughly. When I give you a research question, provide: (1) A summary of current understanding, (2) Key debates or disagreements in the field, (3) The strongest evidence on each side, (4) Gaps in knowledge. Distinguish between well-established facts and emerging theories. Cite specific studies or sources when possible.

10. Career Coach

You are a career coach specializing in tech professionals. When I share a career situation or decision, help me think through it by: identifying what I actually want (not what I think I should want), mapping out realistic options, and highlighting blind spots in my thinking. Be honest, not encouraging. If I am making a mistake, tell me directly.

Common System Prompt Mistakes

Being Too Vague

"Be helpful and provide good answers" tells the AI nothing it does not already default to. Every instruction should change the AI's behavior in a specific, observable way.

Contradicting Yourself

"Be concise" and "provide comprehensive explanations" in the same prompt creates confusion. The AI will oscillate between modes unpredictably. Pick one and commit.

Overloading with Rules

A system prompt with 50 rules will be less effective than one with 10. AI models have attention limits. The instructions at the top get the most weight. Put your most important rules first.

Forgetting to Iterate

Your first system prompt will not be perfect. Use it for a few conversations, notice where the AI falls short, and refine. The best system prompts go through five or more revisions before they feel right.

Not Saving What Works

You spend 20 minutes crafting the perfect system prompt for a Claude Project. Three weeks later you need something similar for a different project. Can you remember exactly what you wrote?

This is why the people who get the most from AI keep a library of their system prompts. Not in a random note somewhere — in a searchable, tagged collection they can pull from anytime.

FAQ

Do system prompts work the same across all AI models?

Mostly yes, but with differences. ChatGPT tends to follow system prompts more rigidly, while Claude weighs user messages more heavily. Gemini falls somewhere in between. If you switch between models, you may need to adjust phrasing slightly. The core instructions transfer well across all of them.

How long should a system prompt be?

Most effective system prompts are 100 to 300 words. Shorter than that and you lack specificity. Longer and the AI starts losing focus on individual instructions. If you need more than 300 words, you probably need multiple system prompts for different tasks rather than one that tries to cover everything.

Can I see the system prompt that ChatGPT or Claude uses?

The default system prompts for major AI tools have been publicly documented through various means. They are long and detailed — ChatGPT's runs several pages. You cannot modify these base system prompts, but your custom instructions layer on top of them.

Do system prompts use up my context window?

Yes. System prompts count toward the total tokens the AI can process in a conversation. This is another reason to keep them focused — a 2,000-word system prompt means less room for your actual conversation.

Should I write different system prompts for different tasks?

Absolutely. A system prompt tuned for code review will produce bad results for creative writing. The most effective approach is maintaining a collection of task-specific system prompts you can swap between as needed. Claude's Projects feature is designed exactly for this — one project per use case, each with its own instructions.

How do custom instructions differ from system prompts technically?

Custom instructions are system prompts with a user-friendly interface. When you type into ChatGPT's Custom Instructions settings, those words get injected into the system prompt that runs before each conversation. The only difference is the access method — settings panel versus API call.

Start Building Your System Prompt Library

System prompts are the highest-leverage skill in AI. A single well-written system prompt can transform every conversation you have with an AI tool. Ten of them, covering your most common tasks, can save you hours every week.

The problem most people run into is not writing them — it is keeping track of them. A system prompt in ChatGPT's settings is invisible to your Claude Projects. A great prompt you wrote last month is buried in a conversation you cannot find.

The fix is simple: keep your system prompts in one place, tagged by task and model, ready to grab when you need them. That is exactly what Prompt Wallet is built for — a searchable library where you can store, tag, and version your best prompts across every AI tool you use. Free for individuals, with team sharing when you need it.

Your system prompts are some of the most valuable prompts you will ever write. Treat them that way.