System prompt
The instructions the operator gives the model BEFORE the customer's message — invisible to the customer. Where you tell Roost 'never recommend a product; always offer a fact-find call'.
A system-level configuration provided to the LLM prior to the user's input that defines the model's persona, operational constraints, and behavioral boundaries. It acts as the primary directive that remains active throughout the session, ensuring the model adheres to specific business rules, such as prioritizing discovery calls over direct sales.
A foundational set of instructions injected into the model's context window during the initialization phase, preceding the user-message sequence. It serves as the system-level directive that scopes the model's role, tone, and forbidden actions, functioning as a persistent constraint layer that is invisible to the end-user but governs all subsequent inference outputs.
evolution
- 2020-06 · historyGPT-3 Release
OpenAI introduces the 'prompt' paradigm where natural language instructions guide model behavior without fine-tuning.
- 2022-11 · historyChatGPT Launch
The introduction of the 'system message' API parameter allows developers to define persistent behavioral constraints separate from user input.
- 2023-03 · historyChat Completions API
OpenAI formalizes the 'system', 'user', and 'assistant' role structure, standardizing how system prompts are handled in production.
- 2023-09 · historyPrompt Injection Awareness
The industry recognizes system prompts as a critical security layer, leading to widespread research on 'jailbreaking' and prompt defense strategies.