Skip to main content
Agent instructions are free-form text that gets prepended to the conversation’s system prompt. There’s no length limit, but the model still has a context window. Keep instructions tight.

​
Anatomy of a good agent prompt

1

Persona, who is the agent

One sentence: “You are the Product Agent for Custral, focused on product strategy and feedback.”
2

Voice. How should it speak

Two or three lines. Examples: “Be concise. Use bullet points for any list of three or more. Don’t apologize.”
3

Constraints. What it must not do

“Never write directly to the database. Always confirm a destructive change with the user first.”
4

Procedures, repeatable patterns

Step lists for common asks. “When the user asks for a feature summary, fetch the feature record, list its open conversations, and produce a one-paragraph summary tied to revenue impact.”

​
What instructions do not control

  • Available tools. Set with the agent’s allowed-tools list and web-access toggle, not in prose.
  • Scope. Set with the agent’s scoped objects and pages. A prompt instruction that says “only read the deals object” is unreliable; use the scope settings instead.
  • Model choice. Set with the agent’s model setting, not in prose.