A system prompt is a set of instructions that tells an AI how to behave before the user starts chatting. The best system prompts follow a 5-step structure: define the role, set the context, specify the task format, add constraints, and include examples. A well-written system prompt can transform a generic AI chatbot into a specialized team member that consistently delivers the output you need.
Three months ago, one of our users, told me her AI assistant was “useless.” She’d describe what she needed, and it would give her generic, Wikipedia-style answers that she could never send to a client.
I asked to see her setup. She had no system prompt. Her AI was starting every conversation as a blank slate. That’s like hiring someone and never telling them what their job is.
We spent 10 minutes writing a system prompt together. Her exact words after the first test: “It’s like a completely different tool.”
It is. And I’ll show you exactly how to do it.
What Is a System Prompt?
A system prompt is a set of hidden instructions that runs before every conversation with an AI. The user never sees it – it runs in the background. Think of it as the job description you give a new hire on day one.
Without a system prompt, an AI model is a generalist. It tries to be helpful about everything, which means it’s not particularly great at anything specific. A system prompt turns that generalist into a specialist.
Example of the difference:
Without system prompt: User: “Write me a follow-up email to a client” AI: “Dear [Client Name], I hope this email finds you well. I wanted to follow up regarding our recent discussion…”
With system prompt: User: “Write me a follow-up email to a client” AI: “Hi Sarah – quick note after our call yesterday. I’ve attached the competitive analysis we discussed. Two things jumped out that I think are worth exploring on our next call…”
Same question. Completely different output. That’s the power of a system prompt.
The 5-Step System Prompt Framework
Step 1: Define the Role
Tell the AI exactly who it is. The more specific, the better.
Weak: “You are a helpful assistant.” Strong: “You are a senior marketing strategist with 10 years of experience working with B2B SaaS companies. You specialize in content strategy and growth marketing.”
The role shapes everything that follows. An AI told it’s a “financial analyst” will use different language, make different assumptions, and provide different levels of detail than one told it’s a “customer support agent.”
Template: You are a [specific role] with expertise in [specific domain]. You work for [context about the business/user]. Your specialty is [what they do best].
Step 2: Set the Context
Give the AI the background information it needs to do its job well. This is where you “teach it about your business.”
Include: Who your clients/customers are Your industry and market Your company’s tone and values Any specific knowledge it should reference Example: You work for a boutique consulting firm that serves mid-market SaaS companies ($5M-$50M ARR). Our clients are typically VPs of Marketing or CMOs who need help with positioning and go-to-market strategy. We’re known for being direct, data-driven, and no-BS.
Step 3: Specify the Task Format
Tell the AI what kinds of tasks it will handle and what the output should look like.
Weak: “Help me with writing.” Strong: “Your primary tasks are: (1) Writing client proposal first drafts, (2) Creating competitive analysis documents, (3) Drafting follow-up emails after client calls. For proposals, use this structure: Executive Summary, Problem Statement, Proposed Approach, Timeline, Investment. Keep proposals under 1,500 words.”
The more specific you are about format, the less editing you’ll do later.
Step 4: Add Constraints
Constraints are the guardrails that keep the AI on track. These are the “don’t do this” rules.
Examples of good constraints:
Never use the phrases “I hope this email finds you well” or “Let me know if you have any questions” Always use UK English spelling (organisation, not organization) Keep all emails under 200 words Never make up statistics. If you don’t have real data, say so Always end proposals with a specific next step, not a vague “let’s discuss” Don’t use jargon like “synergy,” “leverage,” or “best-in-class”
Step 5: Include Examples
This is the step most people skip, and it’s the most powerful. Show the AI exactly what good output looks like.
Template: Here’s an example of a follow-up email in our style:
“Hi [Name], great call today. Three things stood out: [Specific insight from the call] [Action item discussed] [Next step with date] I’ll send over the proposal by Thursday. Let me know if anything changes before then. All The Best, [Name]”
Match this tone and structure for all client emails.
Putting It All Together: A Real System Prompt
Here’s a complete system prompt for a client research assistant (this is similar to what many LaunchLemonade users create):
ROLE: You are a senior business research analyst specializing in competitive intelligence for marketing agencies.
CONTEXT: You work for a digital marketing agency that serves B2B SaaS companies. Our clients need competitive analysis, market research, and industry trend reports. Our audience is marketing leaders (VPs, CMOs, Directors) who are smart, busy, and want insights – not summaries.
TASKS: Competitive analysis reports (company overview, product comparison, positioning analysis, SWOT)
- Industry trend briefings (key trends, implications for our clients, recommended actions)
- Client research briefs (company background, key stakeholders, recent news, potential pain points) FORMAT: Use bullet points and headers for scannability
- Lead with the “so what” – why does this matter?
- Include specific data points and source references where possible
- Keep reports to 500-800 words unless asked for more detail CONSTRAINTS: Never fabricate statistics or data
- If information is uncertain, say “Based on available data” or “This requires verification”
- Use professional but direct language. No fluff, no corporate jargon
- Always end with “Recommended Next Steps” section
- Use UK English spelling
Common System Prompt Mistakes
Too vague: “Be helpful and professional” tells the AI nothing useful. Be specific. 2. Too long: Keep system prompts under 500 words. Beyond that, models start losing focus on the early instructions. 3. Contradictory rules: Saying “be concise” AND “provide comprehensive detail” confuses the model. 4. No examples: Examples are worth more than 100 words of instruction. Always include at least one. 5. Trying to do everything: One role per agent. Don’t make your “email writer” also do “financial analysis.”
Frequently Asked Questions
What is the maximum length for an AI system prompt?
Technically, modern AI models can process system prompts spanning hundreds of thousands of words, effectively allowing you to treat an entire library as a set of instructions. However, just because you can doesn’t mean you should. For peak performance, the ‘sweet spot’ remains 200 to 500 words. While models have better long-term memory than ever, extremely dense prompts can still lead to ‘instruction drift’ or increased latency. To get the best results, prioritize clear, high-impact instructions and use ‘few-shot’ examples rather than exhaustive explanations.
Can I change a system prompt after creating an AI assistant?
Yes. System prompts can be updated at any time. On most platforms including LaunchLemonade, you can edit the system prompt in your agent’s settings. Changes take effect immediately on the next conversation.
What’s the difference between a system prompt and a user prompt?
A system prompt sets the AI’s behavior and personality before any conversation starts – the user typically never sees it. A user prompt is the actual message a person types in the chat. The system prompt shapes how the AI responds to every user prompt.
Do system prompts work with all AI models?
All major AI models (GPT-4, Claude, Gemini, Llama, Mistral) support system prompts, though the syntax varies slightly between providers. On LaunchLemonade, system prompts work consistently across all 21+ supported models – you write it once and it works everywhere.
How do I test if my system prompt is working well?
Send 5-10 different test messages that cover your main use cases. Check: Does the AI stay in character? Does it follow your format rules? Does the tone match your examples? If it drifts on any of these, tighten the relevant section of your system prompt.
Stop spending hours on tasks your AI team member could handle in minutes. Write a great system prompt, and everything changes.
Create your first AI team member with a built-in system prompt builder: https://launchlemonade.app?utm_source=blog&utm_medium=organic&utm_campaign=how-to-write-ai-system-prompts
Cien Solon is the founder and CEO of LaunchLemonade, building AI team members for every business. Follow her on LinkedIn for daily insights on AI, delegation, and building in public.
