Back to glossary

System instruction

Sistem talimatı

D3

A system instruction is a high-level set of instructions and context supplied by an application to steer a chat model’s responses.

Review status: 2026-11-26

Technical explanation

In chat interfaces that support this role, it can specify the assistant’s job, boundaries, tone, output format, and safety guidance before or alongside user input.

Conceptual boundaries

A system instruction is not a model-level safety guarantee and is not interchangeable with an ordinary user prompt, even though both are inference-time input.

Provider-neutral example

A knowledge-base application can set a system instruction to answer only from supplied documents and identify missing evidence.

Limitations

System instructions can conflict with other input or be targeted by prompt-injection attacks, so they need testing and additional controls for consequential use.

Related concepts

Atomic claims and evidence

  1. 1.1Microsoft Learn, System message design for Azure OpenAI
    Source
    Microsoft Learn, System message design for Azure OpenAI
    Source role
    Authoritative source
    Exact locator
    What is a system message?; How system messages work
    Supported claim
    Microsoft defines an Azure OpenAI system message as instructions and context sent to a chat model to guide its responses, including role, boundaries, tone, formats, and safety constraints.
    Last verification
    Review due
    Scope limitation
    This is provider-specific interface guidance for Azure OpenAI and is not a universal cross-provider message hierarchy.
  2. 2.1NIST AI 100-2e2025, Adversarial Machine Learning: A Taxonomy and Terminology of Attacks and Mitigations
    Source
    NIST AI 100-2e2025, Adversarial Machine Learning: A Taxonomy and Terminology of Attacks and Mitigations
    Source role
    Authoritative source
    Exact locator
    Section 3.1.1 printed p. 38: In-context instructions and system prompts
    Supported claim
    NIST describes system prompts as in-context instructions often prepended to other model input, and notes that prompt injection can override intended instructions.
    Last verification
    Review due
    Scope limitation
    This describes a common LLM-application security risk and does not establish that every system instruction will be overridden.