System Prompts

Disclaimer. Provider-specific notes (Anthropic, OpenAI system-message handling) reflect current published behavior. Provider docs remain authoritative for role-precedence semantics and prompt-caching mechanics.

A system prompt is the persistent backstage instruction that shapes behavior across every turn of a conversation. It is not a one-shot role assignment in a user message — it is the operating constitution of the agent.

Use this skill when you are building something that will be used more than once: a product feature, an internal agent, a customer-facing assistant, a CLI tool with an LLM core. Use prompt-engineering-role instead when you want to focus a single ad-hoc message on a particular expert lens.

When to use

• Building a chatbot, agent, or LLM-backed product feature.
• Need consistent tone, format, and scope across long sessions.
• Need explicit boundaries (what the assistant must refuse, deflect, or escalate).
• Multiple developers will extend the assistant's capabilities later.

The Five-Part Structure

Every production system prompt should answer five questions, in order: