> ## Documentation Index
> Fetch the complete documentation index at: https://docs.open.cx/llms.txt
> Use this file to discover all available pages before exploring further.

# AI Instructions

> Write the rules your AI follows and the answers it gives — a system prompt plus versioned training scenarios, all edited in one dashboard.

**AI Instructions** is the dashboard surface where you teach the AI directly — no connectors, no source systems. You write two things here: the **AI Profile** (how the agent behaves) and **training scenarios** (how it answers specific questions). Everything drafts first, publishes when you're ready.

<Tip>
  Open [**AI Instructions**](https://platform.open.cx/ai-instructions) in your OpenCX dashboard.
</Tip>

## What lives here

<AccordionGroup>
  <Accordion title="AI Profile — the agent's persona and ground rules" icon="id-badge">
    A single system prompt that defines who the agent is, the tone it uses, and the rules it never breaks. It's injected into every conversation before retrieval runs. Edit it as rich text; the editor keeps a **draft** alongside the published version so you can revise without affecting live conversations until you hit **Publish**.
  </Accordion>

  <Accordion title="Training scenarios — specific things to say or do" icon="list-check">
    Discrete instructions the AI applies on top of the profile. Each scenario has a type: **Behavioral** (always active — tone, formatting, policy reminders) or **Scenario-specific** (triggered when the customer's question matches). Scenarios carry their own draft/published lifecycle and version history. See [Training scenarios](/ai-instructions/training-scenarios).
  </Accordion>

  <Accordion title="Directories — folders for your scenarios" icon="folder-tree">
    Group scenarios by team, product area, or channel. Directories are flat org-wide — no nesting. Moving a scenario between directories is a drag; visibility settings (channels, segments) can be applied to a whole directory at once.
  </Accordion>

  <Accordion title="Version history per scenario" icon="clock-rotate-left">
    Every publish of a training scenario writes an immutable version. Open the history panel on any scenario to see who published what and when, and roll back — a rollback populates the draft with the old content so you can review before republishing. The AI Profile is a single mutable document and does not keep version history.
  </Accordion>
</AccordionGroup>

## When to use AI Instructions vs. the other pillars

Pick the surface closest to where your content already lives. Don't retype things you can sync.

| Pick this when…                                                                                                          | Use                                                       |
| ------------------------------------------------------------------------------------------------------------------------ | --------------------------------------------------------- |
| The rule is **universal** (tone, never reveal pricing, always confirm order ID)                                          | **AI Instructions** — AI Profile or a Behavioral scenario |
| You have a **specific question** and a **specific answer** that's too important to rely on retrieval                     | **AI Instructions** — Scenario-specific                   |
| The content is a **customer-facing article** you'd want on a public Help Center anyway                                   | [**Help Center**](/help-center/overview)                  |
| The content already lives in **Zendesk, Notion, Confluence, GitBook, Shopify, Intercom, Front, Freshdesk, or a website** | [**Knowledge Sources**](/knowledge/sources/overview)      |

## Draft, publish, and what the AI sees

Every document in AI Instructions — the profile and every scenario — has a **published** state and a **draft** state.

* The draft is what you see while editing. It is **never** served to a customer.
* Publishing copies the draft over the live version. For scenarios, a new immutable version is written so you can roll back later.
* Discarding a draft wipes your unsaved changes and reverts to whatever is currently live.

Because everything is org-wide, a publish takes effect on the next conversation across every channel.

***

## Related Documentation

<CardGroup cols={2}>
  <Card title="Training scenarios" icon="rabbit" href="/ai-instructions/training-scenarios">
    Behavioral vs. scenario-specific, channel and segment targeting, version history.
  </Card>

  <Card title="Help Center" icon="book-open" href="/help-center/overview">
    Author customer-facing articles that also train the AI.
  </Card>

  <Card title="Knowledge Sources" icon="brain" href="/knowledge/sources/overview">
    Sync content from Zendesk, Notion, Confluence, and more.
  </Card>

  <Card title="Agent Knowledge overview" icon="book" href="/knowledge/introduction">
    The three ways OpenCX learns what to say.
  </Card>
</CardGroup>