System Prompt
Every section of the instructions your AI agent reads before it replies, in order: where each comes from, and the expert sections you can edit directly.
Before every reply, your AI agent reads its system prompt: a fixed sequence of sections that together say who it is, how to behave and what it knows about this conversation. Settings -> Automation -> System Prompt lists every section in that order and shows where each one comes from.
Most sections come from other settings or are built in, so this page is mainly for understanding why the AI agent behaves the way it does. The sections you can edit here start from tuned defaults and rarely need changing. Try the writing style, bot rules and workflows first.
For how the AI agent uses its prompt and tools, see the Bot Operating Model.
The sections, in order
| # | Section | Comes from |
|---|---|---|
| 1 | Task | Edited here |
| 2 | Time & Channel | Built in: conversation start, current time and the customer's channel |
| 3 | Safety | Edited here |
| 4 | Internal confidentiality | Built in |
| 5 | Instruction following | Edited here |
| 6 | Tool calling | Edited here |
| 7 | Customer attachments | Built in: how to read images and files |
| 8 | How to write | Writing style, plus output syntax edited here |
| 9 | Business description | Edited here |
| 10 | Additional rules | Bot rules |
| 11 | Workflows | Workflows |
| 12 | Articles | Knowledge base |
| 13 | Handoff | Handoff settings, when the AI agent keeps answering after a handoff |
| 14 | Social media page | Built in, for Messenger and Instagram DMs |
| 15 | Customer Email | Built in, when known |
| 16 | Previous conversations | Built in, when the customer's email is known |
| 17 | Customer Phone | Built in, when known |
| 18 | Browser History | Built in, for Octocom web chat |
| 19 | Recently viewed products | Built in, when product sync is set up |
| 20 | Products | Built in, from the synced catalog |
| 21 | Custom instructions | Your Python prompt section |
Sections with nothing to say for a conversation are left out. For example, Additional rules is missing when the business has no bot rules for the customer's channel.
The sections you can edit here
Each one shows Default or Customized. Your text replaces the default; it doesn't add to it. For Safety and Instruction following that means you are now responsible for the guardrails the default provided, so carry over what you still need. Reset to default brings the default back.
Task
Who the AI agent is: a customer service agent for your business, and how it answers whether it is an AI. The default names your business, or leaves it unnamed when the business is anonymized. Your own text does neither, so only replace it when you need to rewrite that foundation.
Safety
The guardrails it always keeps: staying on topic and not making commitments the business hasn't approved.
Instruction following
Which sources of truth win, for example that workflow instructions outrank articles.
Tool calling
How and when the AI agent uses tools. The default adapts to your live workflows on the customer's channel (embedded or loaded on demand) and is left out when there are none. Your own text is used on every channel and replaces that adaptation, so a bad override can stop the AI agent from calling tools correctly.
Output syntax
The formatting rules for replies: markdown, no raw HTML, no tables. Web chats that show widget elements also get the chat window rules, which win where the two disagree. Change it only when your channel has specific rendering limits.
Business description
A short overview of your company, as context for every reply. It's empty by default and left out until you write one. Keep it an overview: the AI agent searches your catalog and knowledge base on demand, so a product list here only makes every prompt longer.
Custom instructions (Python)
Python you write once per business. It runs on every reply and adds its result as the last section. Editing it needs permission to edit automations. See Custom System Prompt Sections.
See the prompt behind a reply
The settings show how the prompt is built for a business. To see the exact prompt a reply used, with every section filled in for that conversation, ask Copilot to show the system prompt of that message (get_response_system_prompt).
Managing
- In the dashboard: Settings -> Automation -> System Prompt.
- With Copilot or MCP:
get_system_promptreturns every section in order with itssource, the text of the editable sections andsystemPromptSectionCode.update_system_promptpatches them: send a text to override a section,nullto reset it, or leave it out to keep it. Code supports targeted edits. - With the REST API:
GETandPATCH /rest/v1/system-prompt, with the same fields. SendingsystemPromptSectionCodeneeds theautomations:writescope.
| Field | Section |
|---|---|
task | Task |
safety | Safety |
instructionFollowing | Instruction following |
toolCalling | Tool calling |
outputSyntax | How to write: output syntax |
businessDescription | Business description |
systemPromptSectionCode | Custom instructions (Python) |
Reads return the text the AI agent uses, and customizedFields lists the sections your business set. Every change is recorded in the configuration history.