Write the prompt

The User response prompt is the Agent's standing instructions. They go into every message, before the history and the passages retrieved from Collections.

The editor has formatting, a side-by-side preview and a Prompt Library with reusable texts.

What the prompt decides

Standing instructions are good for four things:

• Role: who the Agent is and whom it answers.
• Scope: what is inside and what is outside its subject.
• Format: the length, structure and language of answers.
• Behavior when unsure: what to do when the answer isn't in the material.

The form's default text already covers format and behavior when unsure. It is usually worth adding role and scope, which are specific to your case.

What not to include

• Knowledge: facts, tables and policies belong in a Collection, not in the prompt. Content in the prompt is paid for on every message and can't be updated without editing the Agent again.
• Secrets: keys and passwords never go in the prompt. Access to systems is handled through Credentials and tools.
• Rules that are already guardrails: sensitive-data filters apply to the whole organization and don't depend on each Agent remembering them.

Writing for an Agent with Collections

When the Agent has knowledge attached, the prompt should say what to do with it.

Two instructions handle most cases: answer from the retrieved material, and say it doesn't know when the material doesn't cover the question. Without the second, the model tends to fill the gap with what it learned in training, which produces plausible and wrong answers.

Adjusting later

The fastest way to evaluate a prompt is to chat with the Agent about real cases, not ideal ones. Ambiguous and out-of-scope questions reveal more than well-phrased ones.

Change one thing at a time. Rewriting the whole prompt between two tests makes it impossible to know what improved.

Next steps

• Attach the knowledge the prompt refers to in Collections.
• Put the Agent in people's hands in Publish and maintain.