
System prompt: the standing instructions behind an AI app
In short: A system prompt is the standing set of instructions an AI app sends to the model before any user message, and it guides behaviour without enforcing it.
A system prompt is the standing set of instructions an AI app sends to a language model before the user's messages, defining its role, tone, rules, available tools and output format. The user usually never sees it. OWASP's 2025 Top 10 for LLM applications lists system prompt leakage as risk LLM07 and says the prompt should not be treated as a secret or as a security control.
System prompt in plain words
It is the briefing a call-centre agent gets before their shift: who you are, what you may promise, what to escalate. Customers never read it, but they hear its effects in every call.
How it works
APIs accept the system prompt as a separate field or message role. Anthropic's prompting guide, read in November 2026, says setting a role in the system prompt focuses the model's behaviour and tone, and that even a single sentence makes a difference.
Treat it as strong guidance, because nothing forces the model to obey it. In agents, the harness often assembles the system prompt dynamically from fixed instructions, tool descriptions, memory and project files.
Example
You are the billing assistant for Acme Cloud.
Answer only questions about invoices, plans and payment methods.
Use get_invoice and get_plan tools; never guess amounts.
If the user asks for a refund, collect the invoice ID and call escalate_refund.
Reply in under 120 words. Use INR for Indian customers.
Where you'll meet it
- AI products and custom assistant configurations
- Agent harnesses, where it includes tool and project instructions
- System prompt leakage, risk LLM07 in OWASP's 2025 list
- Prompt-injection attacks that try to override it
What it changes for you
For product teams, the system prompt is a product spec in prose: version it, test it and review changes like code. For developers, keep it focused on behaviour and push facts into retrieval or tools so it stays short. For security teams, assume it can be extracted. OWASP says credentials and connection strings should not be in it.
System prompt vs user prompt
The system prompt is fixed by the builder and applies to every conversation. The user prompt is what each user types, one message at a time.
Common mistake
My view: using the system prompt as the security boundary ("never reveal customer data"). Users and injected content can talk the model around instructions. Enforce access in the tools and data layer, where words cannot change it.
Related terms
Prompt engineering · Prompt injection · Jailbreak · Context engineering · Guardrails
From the archive: Delete a prompt rule only after naming what replaced it
Write behaviour into the system prompt and enforce access in the tools and data layer.


