# System prompt vs guardrails

System prompts ask a model to behave, and the model may not comply. Guardrails are separate checks on each input, output and tool call that enforce a policy from outside the model.

## What each one does

A system prompt is the developer's instruction at the start of the model's context: its role, tone and rules. The model reads it as text, along with everything else.

A [guardrail](/glossary/ai-guardrails) runs outside the model. It reads an input, an output or a tool call, judges it against a [policy](/glossary/guardrail-policy), and acts: block, rewrite, flag.

| | System prompt | Guardrail |
| --- | --- | --- |
| What it is | An instruction | A check with an action |
| Who decides | The model being instructed | A separate checker |
| Under attack | The attacker talks to the same model | The attacker must also get past a second check |
| Record | None | A verdict per request |

## Why instructions fail under injection

A model reads the system prompt, the user's message and any retrieved document as one stream of text. Nothing in that stream marks the developer's lines as binding, so later text that reads like an instruction can win. That is [prompt injection](/glossary/prompt-injection), and a [jailbreak](/glossary/jailbreak) does the same to the model's safety rules. Stronger wording does not change that.

## Which to pick

Both. The system prompt tells the model what to do: its role, its style, the normal path. Guardrails are for what must not happen even when the model is fooled or wrong, such as leaks, unsafe content and harmful tool calls.

## With White Circle

[Control](/products/control) sits between your model and your users. Its policies are written in plain language and enforced on every input, output and tool call.

