Stockertim Dictionary · Artificial intelligence
System prompt
Also called system message, system instructions, custom instructions
In short
A system prompt is a set of standing instructions placed before a conversation starts, which shape every later response, such as rules on tone, format or what the assistant should do.
- Beginner
- Concept
- Checked
In detail
Most AI assistants process two kinds of instruction. The system prompt is set once, usually by whoever builds or configures the tool, and applies to every turn. User messages are the one-off requests that follow. A system prompt typically describes the assistant's purpose, the audience, the tone, the formats to use and the topics to decline or hand to a person.
Because it is always present, a system prompt is the natural place for rules that should never be forgotten, such as quoting sources or never giving personal financial advice. Many tools let individuals add their own standing instructions too, for example a preferred reading level or spelling convention. These are applied alongside any instructions set by the organisation.
A system prompt steers behaviour but does not enforce it absolutely. Long conversations, conflicting user requests or crafted text can push a model away from its standing rules, and the text of a system prompt can sometimes be revealed by users. Anything that must hold for safety or security is backed by checks outside the model, not by the system prompt alone.
Worked example
Kite & Co Accountants sets up an internal assistant for answering staff questions about expenses.
- Weak system prompt: "You are helpful."
- Improved system prompt: "You answer staff questions about the expenses policy pasted below. Quote the policy section for every answer. If the policy does not cover the question, say so and suggest asking the finance team. Plain English, under 150 words."
- Reason: the improved version sets the source, the citation rule, the fallback and the format once, so every later answer follows them.
Answers become consistent, sourced and short, and gaps in the policy are sent to a person.
In practice
Teams write and version system prompts for shared assistants, test them against tricky questions, avoid putting secrets in them, and pair them with outside checks for rules that must not fail.
Watch out
- A system prompt is not hidden or secure; assume users may see it.
- Standing instructions guide the model but are not a substitute for permissions and access controls.
Learn it properly
Official sources
- Top 10 for large language model applications · OWASP Foundation
- Guidelines for secure AI system development · National Cyber Security Centre
Facts last checked on 3 November 2026.
Education, not advice: the Stockertim Dictionary explains what terms mean. It never recommends buying or selling anything.