System prompt
The standing instruction that defines a model’s role, rules and boundaries before any user input arrives. In agent systems it carries identity, tool policy and safety rails — and deserves version control.
The standing contract
The system prompt establishes identity, rules and boundaries before any user input arrives. In an agent it carries more weight than in a chatbot, because it also governs tool policy, escalation behaviour and what the agent must refuse — decisions with real consequences rather than conversational ones.
Treat it as code
It changes behaviour across every run, so it deserves what code gets: version control, review, and a note explaining why each rule exists. The alternative is the familiar organisational artefact — a sprawling prompt full of rules whose origins nobody remembers and which everybody is therefore afraid to touch.
Keep it about identity, not tasks
Standing instructions belong here; task-specific procedure belongs in a skill loaded when relevant. Conflating the two produces a permanently enormous prompt that costs tokens on every call and buries the rules that always matter under ones that rarely do.
In AGNT
AGNT agents keep system prompts as editable, versioned configuration — not text lost in a chat box.
System prompt — common questions
How long should a system prompt be?
Long enough to cover identity, standards and boundaries, and no longer. Push task procedure into skills so the standing cost stays small.
What if instructions conflict?
State an explicit precedence order. Models resolve conflicts arbitrarily otherwise, which produces behaviour that changes between runs for no visible reason.
Can each AGNT agent have its own?
Yes. Every agent keeps its own versioned system prompt as editable configuration rather than text lost in a chat box.