
A system prompt is the backstage instruction that runs before every conversation. Write it well and the AI behaves exactly how you want, every time, without repeating yourself.
TLDR
A good system prompt defines who the AI is, what it is for, how it should respond, and what it should never do. Set it once and it shapes the entire conversation.
Start with the role and purpose
Open with a clear role and reason for that role: "You are a customer support agent for [Brand], helping users resolve billing and account issues." This anchors every response that follows.
Define the output format and style
Be explicit about how responses should look. Short or long? Bullet points or prose? Formal or conversational? Markdown or plain text? If you do not define format, the model will guess differently every time.
List what it should always do
Add two or three consistent behaviors: "Always ask a clarifying question before giving advice," "Always cite your reasoning," "Always end with a follow-up question." These create reliable, repeatable behavior.
List what it should never do
Constraints are as important as permissions. "Never give medical diagnoses," "Never recommend specific financial products," "Never discuss competitors." Explicit prohibitions prevent drift.
Test with edge cases before finalizing
After writing, test with the hardest requests you expect: out-of-scope questions, ambiguous inputs, and the most common failure mode. Revise until the AI behaves correctly on all of them.
Example prompt
A writing coach system prompt: role, behaviors, prohibitions, and format all defined in one block
You are a writing coach helping professional writers improve their craft. Your users are non-fiction authors, journalists, and content strategists with at least two years of writing experience. Your role: - Give direct, specific feedback. Do not soften criticism. - Always explain WHY something is not working, not just that it is not. - When rewriting examples, show the original and rewritten version side by side. - Ask one clarifying question before giving feedback if the writing goal is unclear. Never do: - Praise writing indiscriminately - Give generic advice without applying it specifically to their text - Write full pieces for them Response format: concise, direct, no filler.
Repeated use cases
If you run the same type of conversation regularly, such as customer support, content review, or coaching sessions, a system prompt saves you from re-explaining context every time.
Consistent behavior across a team
When multiple people use the same AI tool, a shared system prompt ensures everyone gets consistent responses with the same persona, format, and constraints.
Building AI-powered products
Every AI chatbot or assistant runs on a system prompt. If you are building with the API, writing a strong system prompt is the single most important thing you will do.
Being vague about output format
If you do not specify format, the model defaults to whatever it thinks is best, which changes based on the question. Always be explicit: length, structure, whether to use bullet points.
Forgetting prohibitions
Most system prompts define what to do but not what to avoid. Adding two or three explicit "never do" rules significantly reduces unwanted responses.
Too many rules
A system prompt with 20 rules is harder to follow than one with 5. Prioritize the behaviors that matter most and iterate from there.
A system prompt is set before the conversation begins and persists across all turns. A regular user prompt is part of the conversation itself. System prompts shape behavior at a foundational level. User prompts direct specific tasks within that behavior.
The concept is the same but the field differs. In ChatGPT it is Custom Instructions or the system field in the API. In Claude it is the system prompt. In Gemini API it is the systemInstruction field. All three respond to the same role-behavior-format-constraints structure.
Long enough to define role, format, key behaviors, and prohibitions: typically 100 to 300 words for most use cases. Longer prompts can cause the model to lose track of early rules, so prioritize the most critical instructions.
Bottom line
A well-written system prompt is a one-time investment that pays off every session. Define the role, the format, what to always do, and what to never do. Test with edge cases and iterate.
Prompt packages that apply this technique directly.
Prompt Engineering
The difference between a useful AI output and a frustrating one is almost always the prompt.
See promptsTechnical Documentation
Creating comprehensive technical documentation can overwhelm teams, leaving critical gaps or unclear instructions.
See promptsContent Marketing Strategy
Most content marketing fails because it produces content without a strategy.
See prompts