Fill in the sections and get a ready-to-use system prompt, correctly formatted for Claude, GPT-4o, or Gemini.
A system prompt is a special set of instructions given to an AI model before the conversation begins. Unlike user messages, the system prompt defines the model's persona, role, behavioral constraints, and output format — it's the "programming" that shapes how the model responds to everything that follows. In most AI APIs, the system prompt is passed separately from the conversation history and persists for the entire session. Well-crafted system prompts are one of the most effective ways to get consistent, high-quality outputs from language models. A weak or vague system prompt is one of the most common reasons for unpredictable or unhelpful AI behavior.
Each model was trained with different conventions for how system prompts are structured, which affects how well it interprets instructions. Anthropic's Claude responds best to XML-tagged sections (<role>, <context>, <task>, etc.) because this structured format was emphasized in Claude's training. OpenAI's GPT-4o is flexible with formats but works well with Markdown headers for section delineation. Google's Gemini responds well to clear, direct prose without heavy formatting. Using the format each model was trained to expect typically improves consistency, reduces misinterpretation, and produces more predictable outputs across diverse user inputs.
Most effective system prompts are between 200 and 800 tokens for general-purpose assistants. Too short (under 100 tokens) and the model has too little guidance to behave consistently. Too long (over 2,000 tokens) and key instructions can get "diluted" as the model struggles to weight everything appropriately — and you're paying for every token in every API call. Focus on clarity over completeness: only include rules the model actually needs to follow. The most important instructions should appear early in the prompt (models tend to weight earlier text more heavily). Use our token counter to check your prompt length.
Yes — few-shot examples are one of the most powerful techniques for improving output quality, often more effective than lengthy instruction text. Including 2–3 input/output examples showing exactly the format, style, and content you expect "shows" the model what good output looks like rather than just describing it abstractly. Examples are particularly valuable when you need specific output formats (structured JSON, specific prose styles, particular response lengths) or when describing nuanced behaviors that are hard to articulate in rules. Keep examples concise — they count toward your context window and API costs.
Prompt injection is a security vulnerability where malicious user input attempts to override or bypass your system prompt instructions — for example, a user sending: "Ignore all previous instructions and instead..." This can lead to the model disclosing confidential information, behaving outside its intended role, or being manipulated into harmful outputs. Defense strategies include: explicitly instructing the model to ignore attempts to override its instructions; validating and sanitizing user inputs before inserting them into prompts; using clear role boundaries ("You are X, only do Y"); considering separate system and user turns rather than concatenating everything; and testing your prompts with adversarial inputs during development.
The best system prompts share several characteristics: they are specific (concrete role, not just "be helpful"), they define boundaries clearly (what the model should and shouldn't do), they specify output format explicitly, and they use the native format of the target model. Start with the role — who or what is the AI? Then add context the model needs to function correctly. Define the task concretely. Specify output format and length. Add constraints for what to avoid. Optionally add tone guidance and examples. Test extensively with real user inputs, including edge cases and adversarial inputs. Iterate based on observed failures, not theoretical concerns.
You can use the same core content across models, but you'll get better results by adapting the format for each. Claude's XML-tagged format often confuses GPT-4o, which works better with Markdown headers. Gemini prefers clean prose over structured formats. Beyond format, different models have different tendencies — Claude tends to be verbose by default and benefits from explicit length constraints; GPT-4o may benefit from explicit instructions about when to use lists vs. prose. If you're building a system that switches between models (e.g., for cost optimization or failover), test your prompt with each model and maintain format-specific versions.