Every AI model you interact with has two types of input: the conversation you see, and a set of instructions you usually don't. Those invisible instructions are the system prompt.
When you chat with a customer service bot, a coding assistant, or a specialized AI tool, there's almost certainly a system prompt shaping every response — setting the persona, constraining topics, defining tone, providing background context about the product or company.
You can write system prompts too. And once you learn how, you'll stop wanting to go back.
What a System Prompt Does
A system prompt is instructions that apply to the entire conversation, not just one message. While a regular prompt says "do this specific thing," a system prompt says "this is how I want you to behave at all times."
Think of it as the difference between:
- User message: "Give me feedback on this email" (one-shot)
- System prompt: "You are a communication coach. Always lead with what's working before critiquing what's not. Prioritize clarity and directness above all else." (standing rules)
The system prompt shapes every response that follows. It's where you establish:
- Persona — Who is the AI in this context?
- Tone — How should it communicate?
- Scope — What should it focus on or avoid?
- Background knowledge — What does it need to know about you or your context?
- Behavioral rules — How should it handle edge cases?
Where System Prompts Live
Different interfaces expose system prompts in different ways:
Claude.ai — Under Settings → "Custom instructions" or in the model's configuration panel. Applies to all conversations in that context.
ChatGPT — Settings → Personalization → "Custom Instructions." Two fields: what the model should know about you, and how you want it to respond.
API access — Passed directly as the system parameter in each API call. Full control.
Specialized tools — Many AI-powered products let you configure a "persona" or "behavior" — that's a system prompt under the hood.
Anatomy of a Good System Prompt
A strong system prompt has four components:
1. Identity
Who is this AI right now?
You are an expert technical writer who specializes in documenting
software products for developer audiences.
2. Behavioral Rules
How should it act consistently?
- Always prioritize accuracy over completeness — if you're not sure, say so
- Use active voice and short sentences
- When reviewing drafts, give specific edits, not general praise
- Never use corporate jargon ("leverage," "synergy," "ecosystem")
3. Background Context
What does it need to know to do the job well?
I work at a developer tools company called Forge. Our main product is a
CI/CD pipeline tool used by engineering teams of 5–50 people. Our users
are developers, not DevOps specialists — they want things to "just work."
4. Output Constraints
What should every response look like?
Default response length: concise. Only go long when the complexity requires it.
Format: prefer prose over bullet lists. Use code blocks for any code.
Never start a response with "Certainly!" or "Great question!"
A Full System Prompt Example
Here's what a complete personal assistant system prompt looks like:
You are my writing assistant and thinking partner.
About me:
I'm a product manager at a B2B SaaS company. I write internal strategy
docs, external blog posts, and a lot of emails. I'm trying to become a
clearer, more direct communicator.
How I want you to work with me:
- Be direct. Don't soften feedback to protect my feelings.
- When editing my writing, show me what you'd cut (not just that it's too long)
- If I ask your opinion, give one — don't hedge into uselessness
- When I'm working through a decision, push back on my reasoning if it's weak
- Don't add caveats about AI limitations unless they're genuinely relevant
What I don't want:
- "Great question!" openers
- Bullet points for everything — use them when it's actually list-shaped content
- Recommendations to consult professionals for every judgment call
- Passive voice
Format: match the format to the content. Short questions get short answers.
This system prompt means every conversation starts with a model that already knows who you are, how you prefer to communicate, and what annoys you. Zero re-setup time.
System Prompts vs. Regular Prompts: When to Use Which
| Situation | Use | |---|---| | Recurring task type you do regularly | System prompt | | Context that changes every conversation | User prompt | | Tone/style preferences that are always the same | System prompt | | Task-specific instructions | User prompt | | Your background and role | System prompt | | The actual thing you want done today | User prompt |
A useful mental model: system prompts are for the stuff you'd tell a new assistant on their first day. User prompts are for the specific tasks you assign each day.
Role-Specific System Prompts
You don't have to have one system prompt. You can build different ones for different modes:
Code review mode:
You are a senior software engineer doing a code review. Be direct about
problems. Categorize issues by severity: Critical (will cause bugs/security issues),
Important (affects maintainability), and Style (minor). Explain why, not just what.
Brainstorm mode:
You are a creative thinking partner. Your job is to generate divergent ideas —
unconventional ones welcome. Don't filter for practicality in the brainstorm phase.
For each idea, add one sentence on why it might be interesting. No disclaimers.
Research mode:
You are a research assistant who values accuracy above completeness.
Always flag when you're less certain about a claim. Distinguish between
established consensus and contested claims. For any specific statistic,
tell me where I should verify it.
Switching between modes takes seconds. The payoff is that each mode is already calibrated for its purpose.
Common System Prompt Mistakes
Overloading it — A 2,000-word system prompt with 50 rules is usually less effective than a focused 200-word one. Prioritize the constraints that matter most.
Contradictory instructions — "Be concise. Always provide thorough context for every recommendation." Pick one. When instructions conflict, model behavior becomes unpredictable.
All don'ts, no dos — "Don't use bullet points, don't be too long, don't be formal" tells the model what to avoid but not what to aim for. Balance prohibitions with positive descriptions of what you want.
Setting it and forgetting it — System prompts benefit from iteration just like regular prompts. If you're still getting something that annoys you, add a rule that addresses it.
Key Takeaways
- System prompts are standing instructions that apply to every message in a conversation
- The four components: identity, behavioral rules, background context, output constraints
- Use system prompts for persistent preferences; user prompts for specific tasks
- Different system prompts for different modes dramatically increases consistency
- Iterate your system prompt the same way you iterate regular prompts
Next: what to do when you need to work with a long document or large amount of context — and the model needs to actually reason about all of it. Prompting With Long Documents →