LLM Prompt Template Builder

The system prompt sets the AI's persona, constraints, and behavior rules — it persists across the conversation. The user message is the per-request input from the user. For most LLM APIs, the system prompt is sent in the 'system' role, while user inputs use the 'user' role. Always separate concerns: put instructions in system, data in user.

Enable JSON mode when you need structured output you'll parse programmatically (extracting entities, classification labels, function arguments). Always include a JSON schema example in your prompt when using JSON mode — models perform much better with a concrete example than abstract instructions.

Be specific and concrete. Define: persona (who the AI is), task (what it does), constraints (what it never does), output format (structure/length), and tone. Use XML-style delimiters (<context>, <question>, <output>) to clearly separate sections. Short, explicit instructions outperform long verbose ones.

Variables ({{variable_name}}) are placeholders replaced at runtime with actual values. Use variables for: user inputs, retrieved context (RAG), dynamic data. Keep static instructions in the system prompt and dynamic data in the user message. This makes your prompt cacheable (Anthropic prompt caching, OpenAI cached completions).