System Prompt
The standing instructions a model receives before any user input — identity, scope, rules, and refusal behavior — the closest thing an AI system has to a job description.
Production AI is a system: context, tools, permissions, traces, evals, and feedback loops around the model.
What it is
The system prompt is the text the operator controls: who the assistant is, what it may and may not do, how it should answer, when it should decline or escalate. The user's message arrives after it and is interpreted in its light.
Why it matters
Most behavior people attribute to 'the model' is actually the system prompt. Two products on the same model feel different because their standing instructions differ. It is also a governed artifact: a wording change can shift refusal behavior, tone, or tool usage across every conversation at once.
How we manage it
System prompts are versioned files in a repository, not strings buried in code. Changes ship through the same gate as model changes: run the eval set, compare against production, promote only when the numbers pass. The persona layer of the prompt is tested separately for stability under adversarial pressure.
Related resources
The versioned, tested set of rules and templates that govern how prompts are assembled for an AI workflow — instructions, examples, formatting, refusal patterns, escalation language.
The discipline of deciding what an AI model sees on every call — instructions, retrieved data, memory, tool definitions, examples — and how to assemble them reliably as the workflow grows.
Before an AI agent can be useful to anyone, it has to be something — a coherent identity that holds up across users, sessions, and adversarial pressure. This is the research track that defines what that means and how to keep it stable.
Side-by-side measurement of a candidate prompt or model against the current production version on the same eval set — the unit of safe change in a serious AI workflow.