System prompts typically cover who the assistant is, what subject matter it may address, what it must refuse, how to format answers, and what to do when it lacks the information to respond. Models generally weight these instructions more heavily than user messages, which is what makes them the primary control surface for behaviour.
They are guidance rather than a hard security boundary. Users can attempt prompt injection to override the instructions, so robust systems pair the system prompt with server-side enforcement, restricted tool access and output checks instead of relying on wording alone.
Clerkzo's guardrails start with a system prompt that keeps the assistant tied to the business it represents, but they do not end there. Off-topic requests are refused, injection attempts are resisted, and questions outside the retrieved knowledge are handed to a human rather than guessed at.
Related terms
Browse all 142 terms- Prompt injectionPrompt injection is an attack where a user types instructions to an AI system, like "ignore your previous instructions," attempting to trick it into abandoning its rules and behaving outside its intended job.
- AI agentAn AI agent is a software system that uses a language model to work towards a goal by deciding what to do next, calling tools or APIs, and acting on the results. It goes beyond generating text to taking actions in a loop until the task is done.
- Conversion rate optimization (CRO)Conversion rate optimization (CRO) is the systematic practice of increasing the percentage of website visitors who complete a desired action, such as submitting a form, booking a call, or making a purchase. It combines behavioural research, hypothesis-driven changes, and measurement to improve results from existing traffic rather than buying more of it.
- WebhookA webhook is an automated HTTP request that one application sends to a URL you control the moment a specific event occurs. Instead of your system repeatedly asking an API whether anything has changed, the source pushes the event data to you as it happens.