Why Structure Matters
Sending a raw question to an LLM works for simple lookups, but falls apart for anything that requires a specific tone, format, or role. A structured prompt separates concerns: the system message defines who the model is and what rules it follows, the user message provides the actual request, and optional assistant prefills steer the opening of the response.
What Each Role Does
- System: Sets persistent context — persona, output format, restrictions, language. This is where "you are a concise JSON API" lives.
- User: The real input. Keep it focused on one task per turn.
- Assistant: A partial response that forces the model to continue in a specific direction (useful for JSON opening brackets or code fences).
Token Estimation Tips
Every token costs money and counts against the context window. Common rules of thumb: one token ≈ 4 English characters, code is denser. Put stable context (system) in a cached prefix when your provider supports prompt caching. Keep dynamic parts (user) short and specific. Aim for the smallest prompt that still gives you the output quality you need — over-specifying wastes tokens without improving results.
Getting Started
Use a prompt builder tool to assemble roles visually, see the estimated token count before sending, and copy the final payload as a provider-ready JSON object.