Instructions is the main place you tell an agent who it is and how to behave — its system prompt, built from a role, a goal, a detailed instructions body, and the conversation starters users see first.
Overview
The section's own description sums it up: "Define who your agent is, how it should behave, and what users see when starting a conversation." You fill three structured fields — a role, a goal, and the long-form Instructions body — and add the conversation starters users see when they first open the agent.
This is distinct from the Prompts section, which holds named, reusable prompt templates for specific functions; Instructions is the agent's core, always-on guidance.
Setting the Agent's Role and Goal
The screen opens with two labeled boxes. Each is a single autosizing line — keep them concrete, since together they frame who the agent is and what it's for, and they anchor everything else it does.
Field | Hint text |
|---|---|
Role | "Persona or expertise of the agent - Ex: Sales Rep." |
Goal | "Describe what this agent is trying to achieve." |
Writing the Instructions Body
Below Role and Goal is the Instructions field — a preview box hinting "Detailed rules for tone, format, and behavior." Click it (or the pencil icon) to open a full-screen editor, headed "Provide background context and instructions related to your agent's expertise."
The editor has a Preview / Code toggle, plus a floating toolbar to Add variable (inserts a {{ }} placeholder) and Add code (inserts a {% %} block); variables show as pills in both the editor and the preview box. Click Done to bring your text back to the section. There's no enforced length limit, but concise, well-ordered instructions produce more reliable behavior than long ones.
Note: The editor is an overlay over the Instructions section, not a separate page — it closes on its own the moment you navigate to another configuration section, including with the browser's back button. Finish with Done before you move on, or you'll come back to find the editor shut.
The Instructions Editor's More Menu
Inside the full-screen editor, a row labeled Add Instructions sits directly above the text itself and carries the more (three-dots) menu, a copy button (once there's content to copy), and the Preview/Code toggle — separate from the editor's outer title bar and Save button.
Group | Actions |
|---|---|
Import / export | Upload instruction file — import a |
Content actions | Copy instruction — copies the text (the item briefly changes to "Copied instruction" to confirm). Clear all — clears the whole field. Discard & close — closes the editor without keeping your changes. |
Note: Copy instruction and Clear all appear only when there's content to act on.
How Instructions Become the Model's System Prompt
At runtime, your role, goal, background, and Instructions body are compiled into the agent's system prompt using a template matched to the model family — different providers get differently structured prompts, so the same instructions work across models. The runtime then appends a few things you don't write by hand:
Language and style directives from your response-generation settings.
Content-filter and word-filter guidance from your guardrail configuration.
Optionally, the output of a custom "additional context" automation (a pre-processing setting) — it runs with the user's query and its result is appended to the compiled system prompt, which is how teams inject live per-request context.
A hidden context message carrying the current date and time (workspace timezone) plus any per-user context, so answers stay time-aware. With prompt caching on, only the date is used so the cache survives all day.
Your instructions are re-compiled fresh on every reasoning step, so edits take effect on the next turn — but when a task/topic is active, its instructions arrive in the conversation (as the task-selection result), layered on top of this system prompt rather than replacing it.
Conversation Starters
Conversation starters are the suggested prompts users see when they first open the agent, so they know what to ask. Add them in the Conversation starters box: each row is a single prompt (placeholder "Start typing…"), and you remove a row with the × that appears on hover. You can add up to four starters.
The box always keeps one empty row ready to type into until all four are filled, so there's no "add row" button to hunt for — and that trailing empty row has no ×, since there's nothing to delete. Pressing Enter in a row jumps to the next one, which makes adding all four a straight run of typing.
Saving Your Changes
Use Save at the top of the Instructions section. While you have edits, an Unsaved changes indicator appears next to it; once saved, the button is disabled because there are no changes to store. Changes you make in the full-screen editor come back to the section when you click Done — they aren't stored until you Save the section.
Note: If the section is read-only — because you're viewing a published version or lack edit access — the Save control won't appear and the fields can't be edited.