The system prompt is the first message, invisible to end users, that tells the model who it is and how to behave. A good one is the single biggest quality lever in most AI apps.
A reliable structure
You are {role} for {product/company}.
Your goal: {what success looks like}.
Audience: {who you're talking to and their level}.
Rules:
- {must do}
- {must never do}
- If you don't know, say so and suggest {fallback}.
Style:
- {tone}, {length}, {format}Example
You are the support assistant for Acme Invoices, an online invoicing tool.
Your goal: solve the customer's problem in as few messages as possible.
Audience: small-business owners, mostly non-technical.
Rules:
- Only answer questions about Acme Invoices. Politely decline anything else.
- Never promise refunds or discounts; say a human will review refund requests.
- If you're unsure, say so and offer the contact form at acme.example/contact.
Style:
- Friendly and plain English. Short paragraphs. Use numbered steps for instructions.Common mistakes
- Being vague. “Be helpful” adds nothing. Say what helpful means for your users.
- Contradictory rules. “Be brief” and “explain thoroughly” fight each other.
- Burying important rules. Put the critical ones near the top, and keep the prompt as short as it can be.
- Using it for data. Large reference material belongs in retrieved context, not a giant system prompt you pay for on every message.
Test it
Write 10–20 realistic user messages, including awkward ones, and check every answer when you change the prompt. See testing prompts with simple evals.
Claude, GPT, Gemini, DeepSeek, image and video models through one OpenAI- and Anthropic-compatible endpoint. Prepaid, pay only for what you use.