Instructions

Instructions are the agent's behavior in your own words: what it is for, who it serves, how it should answer, and what it must not do. They are sent with every request, so they apply to every channel the agent is deployed to.

The Instructions tab#

The tab is an outline on the left and an editor on the right:

• Main instructions holds the agent's core instructions: its role, who it serves, and the rules that apply to every answer.
• Sections holds named blocks of extra guidance. Click Add section to add one, and give each block one subject: the tone to use, an escalation rule, a list of team-specific terms.

The agent reads all of it in every conversation, with the sections after the main instructions, in order. Select an item in the outline to edit it, then click Save or press Command+S on macOS or Ctrl+S elsewhere.

Blocks are easier to revise than a single page, and the name tells you where to look when the agent gets something wrong. A support agent might carry main instructions describing its role, plus Tone, When to escalate, and Product vocabulary sections.

Writing instructions that hold up#

State what the agent should do when it does not know. Without that rule an agent fills gaps with plausible text.

If the handbook does not cover the question, say so and point the user to the People team. Do not infer policy.

Name the audience. "Answer employees" and "answer customers" produce different levels of internal detail from the same knowledge base.

Say what to leave out. Agents default to thorough; most channels reward brevity.

Keep facts that change out of instructions. Prices, staffing, and policies belong in knowledge, where they update on sync instead of requiring an edit here.

Improve with chat#

The Instructions tab has an Improve with chat panel. Describe the behavior you want in plain language and it proposes edits to the blocks, which you review before they are saved.

This is the fastest way to act on a bad answer: paste the answer, say what it should have been, and apply the proposed change.

Related#

• Testing and improving — try changes before they reach users
• Knowledge — facts the agent looks up rather than memorizes
• Agent settings — model, reasoning effort, and access