Prompt Template

A prompt template is a parameterized structure that defines the static instructions, formatting, and message layout of a prompt, with placeholders that are filled with dynamic content at runtime to produce the final prompt sent to an LLM.

Details

Templates separate what the developer authors (instructions, role assignments, output format specifications) from what the system provides at request time (user input, retrieved context, tool outputs, conversation history). This separation enables reuse across requests, structured prompt engineering iteration, and prompt management workflows where templates are versioned and tested independently of application code.

Common template structures range from simple string interpolation (inserting variables into a text block) to message-array templates that define a sequence of role-tagged messages (system, user, assistant) with per-message placeholders. More complex templates include conditional sections that are included or omitted based on available context, and loops that expand over collections such as in-context learning examples or retrieved documents.

Structuring templates with a stable prefix (system instructions, few-shot examples) followed by variable content (user input, retrieved context) improves prompt caching hit rates, since caching operates on token-level prefix matches.

Examples

• A customer support template with a fixed system message defining tone and policy rules, followed by placeholders for the customer's question and relevant knowledge base articles.
• A classification template with few-shot examples in fixed positions and a placeholder for the input text to classify.
• An agent system prompt template with conditional sections for available tools and skills that are included based on the current task.