Most people use AI the same way every session: open a chat, type their request, get output, repeat. Every conversation starts from zero. The AI doesn't know who you are, what you're working on, what tone you prefer, or what format you need.
System prompts fix that. They're the persistent instructions that run before every conversation — the seat adjustment you set once and don't think about again. Without one, you're re-explaining your context every single time. With one, you skip straight to the work.
You don't need to write code to use them. You don't need an API key. You just need to know what to put in them.
What a system prompt actually is
When you open ChatGPT or Claude and start typing, there are two places instructions can come from: the system prompt (set before the conversation starts) and the human turn (what you type in the chat box). The model sees both, but it treats them differently — system instructions carry more weight and persist across the entire conversation.
In practice: if your system prompt says "always respond in bullet points," the model will use bullet points even if you forget to ask. If it says "you are a nutritionist specializing in plant-based diets," that persona holds for every message in that session.
This is the difference between retraining the model (which you can't do without API access) and steering it (which anyone can do).
Where to set a system prompt
You don't need any technical tools. All the major AI products have a place for this:
ChatGPT Custom Instructions: Go to your profile icon → "Customize ChatGPT." You'll see two text boxes — one for "What would you like ChatGPT to know about you?" and one for "How would you like ChatGPT to respond?" The second box is effectively your system prompt. It applies to all chats.
Claude Projects: In Claude.ai, create a new Project (sidebar → "New Project"). Inside the project, there's a "Project Instructions" section at the top of the settings panel. Anything you write there becomes the system prompt for every conversation in that project. Claude Projects also let you upload context files — particularly useful for larger documents.
Gemini Gems: In Google Gemini, you can create a "Gem" (custom AI) with specific instructions. Click the grid icon → "Gem manager" → "New Gem." The instructions field is your system prompt.
API access (for developers): If you're calling models via API, the system prompt goes in the system field of your request. But if you're reading this guide, you probably don't need to think about this yet.
For most people, ChatGPT Custom Instructions or Claude Projects are the right starting points. I'd recommend Claude Projects specifically because they let you create multiple isolated setups — one for marketing work, one for coding, one for personal use — each with its own system prompt.
The five components of an effective system prompt
I've written dozens of system prompts across different use cases, and the ones that work reliably all share the same five components. Skip any one of them and you'll notice the gaps.
1. Role + expertise
Start by establishing who the AI is in this context. Not just a job title — a specific expertise profile.
Weak: "You are a marketing assistant." Strong: "You are a content marketer with 10 years of experience writing for B2B SaaS audiences. You've built editorial calendars, managed content operations, and understand SEO deeply."
The specificity matters. The model draws on different knowledge depending on how you frame the role. "Marketing assistant" produces generic output. "B2B SaaS content marketer who understands SEO" produces output with actual vocabulary and assumptions baked in.
2. Audience + context
Tell the model who it's helping and what that person's situation is.
"You're helping a solo founder who runs a 3-person product company. She has strong domain expertise but limited time. She doesn't need definitions explained — she needs decisions made faster."
This component prevents the model from over-explaining basics to someone who doesn't need them, or skipping context that a beginner would need.
3. Output format
Describe the default format you want for responses. This is where most system prompts leave money on the table.
"Default output format: when I ask for a document, use headers and bullet points. When I ask a direct question, answer in plain prose — no bullets unless the question is a list. Always give the answer first, then explanation if needed."
Without this, the model picks its own format based on how it interprets your question. Often that's fine. Often it isn't.
4. Constraints (what NOT to do)
Explicitly banning the defaults you hate is more effective than describing the positive alternative.
"Do not use phrases like 'certainly,' 'of course,' 'great question,' or 'I'd be happy to.' Do not open with a definition of the topic I asked about. Do not add a summary paragraph at the end unless I ask."
These constraints consistently improve output quality more than almost any positive instruction. The model has strong defaults baked in from RLHF training — many of which are annoying. Banning them directly overrides those defaults.
5. Tone
Describe the register you want, and describe it through contrast.
"Write like a knowledgeable colleague, not a formal report. Conversational but precise. Short sentences when possible. No corporate jargon — if you catch yourself writing 'leverage' or 'synergize,' replace it with a plain verb."
Step-by-step: building a system prompt for a marketing assistant
Let's put the five components together. You want a Claude Project set up for your marketing work — you're a content marketer who writes for a healthcare software company.
Step 1: Define the role
You are a content marketing strategist specializing in B2B healthcare technology. You understand HIPAA compliance at a conceptual level, can write for clinical and non-clinical audiences, and have experience with long-form content, email sequences, and LinkedIn.
Step 2: Add audience + context
You're working with a content marketer at a 50-person health tech company. The company sells EHR software to independent medical practices. The marketer has 5 years of experience and doesn't need basics explained — they need fast, high-quality output they can edit and ship.
Step 3: Specify output format
Output format:
- Blog posts and long-form content: use H2 headers, short paragraphs (3–5 lines max), no fluff intro paragraphs
- Email copy: subject line first, then body, then a note on CTA phrasing
- Bullet lists: only when the content is genuinely list-like; don't convert prose into bullets
- Quick questions: answer first, then context if needed
Step 4: Set constraints
Constraints:
- Do not use corporate filler: "leverage," "streamline," "holistic," "robust," "best-in-class"
- Do not explain what a word or concept means unless asked
- Do not add a "conclusion" or "summary" section unless explicitly requested
- Do not use "certainly," "of course," "great question," or any version of "I'd be happy to help"
- Do not make up statistics or cite sources you're not certain about
Step 5: Set tone
Tone: Direct and informed. Write like a smart practitioner, not a consultant. Use contractions. Be specific rather than general. When you have a recommendation, give it — don't hedge.
The complete system prompt:
You are a content marketing strategist specializing in B2B healthcare technology. You understand HIPAA compliance at a conceptual level, can write for clinical and non-clinical audiences, and have experience with long-form content, email sequences, and LinkedIn.
You're working with a content marketer at a 50-person health tech company selling EHR software to independent medical practices. She has 5 years of experience and doesn't need basics explained.
Output format:
- Blog posts: H2 headers, short paragraphs, no fluff intro
- Email copy: subject line → body → CTA note
- Bullet lists only when content is genuinely list-like
- Quick questions: answer first, context second
Constraints:
- No corporate filler: leverage, streamline, holistic, robust, best-in-class
- No explaining terms unless asked
- No summary sections unless requested
- No "certainly," "of course," "great question"
- No made-up statistics
Tone: Direct, informed, practitioner-voice. Use contractions. Be specific. Give recommendations without hedging.
That's 220 words. Not long. But it fundamentally changes what Claude produces for every message in that project.
Common mistakes that kill system prompt effectiveness
Too vague: "Be helpful and professional." This adds nothing. Every system prompt should be more specific than the model's defaults, not less.
Too long: I've seen system prompts over 2,000 words. Models do read them, but there's a practical limit to how many instructions they'll faithfully track mid-conversation. Keep it under 500 words for conversational use. If you need more, use Claude Projects' document upload feature for the detail and keep the system prompt for the key behavioral rules.
Conflicting instructions: "Be concise" in one line, "Always provide detailed explanations with examples" in another. The model will pick one and you won't know which.
No output format spec: This is the most common gap. Without it, format is left to the model's interpretation. Sometimes that's fine. For regular workflows, define it.
Forgetting to test it: Write the system prompt, then send 5 different types of messages and see if the behavior matches your spec. System prompts often need one round of revision after you see how the model interprets them.
A reusable template
Here's a skeleton you can fill in:
Role: You are a [specific role] with expertise in [domain]. [1–2 sentences of expertise detail.]
Context: You're helping [who] who [situation]. They [do/don't] need [what level of explanation].
Output format:
- [Type of output]: [format spec]
- [Type of output]: [format spec]
Constraints:
- Do not [thing you hate]
- Do not [another thing you hate]
- Do not [a third thing]
Tone: [Describe in contrast — not "professional" but "professional like X, not like Y"]
For real-world examples of what this looks like across different use cases, our system prompts post has a dozen templates you can copy and adapt.
Once you've set up your first system prompt and seen the difference it makes, you'll wonder why you didn't do it sooner. The 20 minutes you spend writing it comes back immediately — every single conversation starts exactly where you want it.
