## Customize conversational AI

URL: https://docs.atlan.com/product/capabilities/atlan-ai/how-tos/customize-conversational-ai

> Write custom instructions to shape how conversational AI answers questions across your organization—set vocabulary, preferred asset types, and response language.

# Customize conversational AI

Custom instructions let you shape how conversational AI answers questions across your organization. Use them to teach conversational AI your vocabulary, state which asset types to prefer or exclude, and set the response language.

Before writing instructions, see [Custom instructions for conversational AI](https://docs.atlan.com/llms/governance/conversational-ai/custom-instructions-conversational-ai/llms.txt) to understand how they work, what patterns to use, and what to avoid.

## Prerequisites

Before you begin, make sure you have:

- **Admin access**: you must be an Atlan admin who can access **Settings** and **Labs**.
- **Conversational AI enabled** for your tenant. See [Set up conversational AI](https://docs.atlan.com/llms/governance/conversational-ai/setup-conversational-ai/llms.txt).

## Set custom instructions

1. In your Atlan workspace, click **Settings** from the left menu.
 * If you are using the **Old UI (Classic)**, go to **Admin settings** from the left navigation panel.
2. Click **Labs** and find the _Atlan AI_ section.
3. For _Customize Conversational AI_, click **Add instructions**.
4. In the _Enhance chat responses_ dialog, for _Custom instructions for chat_, write your instructions—one rule per line.
5. Click **Save**.

Changes apply to conversations started after you save. Open a new chat to test them—an in-progress conversation keeps the instructions it started with.

:::note Conversational AI and enrichment use separate instructions
_Customize Conversational AI_ shapes chat answers. _Customize Atlan AI_, in the same _Labs_ section, shapes generated descriptions, READMEs, and other [enrichment suggestions](https://docs.atlan.com/llms/governance/atlan-ai/what-is-atlan-ai/llms.txt). Setting one doesn't change the other.
:::

## Test your changes

Custom instructions affect every user, so treat a change as a change to production:

1. Write down five to ten questions your users actually ask, and note the current answers.
2. Save your instruction change.
3. Start a new conversation and re-ask the same questions.
4. Compare. Check that the intended rule fired, and that unrelated answers didn't get worse.
5. If answers degraded, remove the most recent instruction and re-test. Introducing rules a few at a time makes it obvious which one caused a regression.

Pay particular attention to questions that worked before your change. The cost of a bad instruction shows up as broad, gradual degradation rather than as an obvious error.

## Next steps

- [Custom instructions for conversational AI](https://docs.atlan.com/llms/governance/conversational-ai/custom-instructions-conversational-ai/llms.txt): how instructions work, what to include, what to avoid, and how they compare to skills, memory, and personas.
- [Troubleshoot custom instructions](https://docs.atlan.com/llms/governance/conversational-ai/custom-instructions-conversational-ai/llms.txt): fix instructions being ignored, degraded answers, and replies without searching.
- [Set up conversational AI](https://docs.atlan.com/llms/governance/conversational-ai/setup-conversational-ai/llms.txt): enable conversational AI and choose who can use it.
- [Conversational AI features](https://docs.atlan.com/llms/governance/conversational-ai/conversational-ai-capabilities/llms.txt): what users can do inside a conversation.

---