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

# Instructions

> Define an AI agent's specialty, communication style, support boundaries, and escalation rules.

Instructions define an AI agent's role, how it speaks to customers, and when it hands over to your team.

When you [create an agent](/product/ai-agents/configuration.md#create-an-agent), describe the problem it should solve. AI uses that description to generate initial instructions. Review these on the setup page and adjust them to match your support policy before assigning tickets.

To make changes later, open the agent under **AI Agents**, edit **Instructions**, and select **Save instructions**. Instructions can contain up to **10,000 characters**. Saved changes apply from the agent's next turn.

## What to include

* **Specialty:** which requests the agent should handle.
* **Communication style:** tone, level of detail, and how to ask for missing information.
* **Boundaries:** what it should avoid promising or attempting.
* **Escalation rules:** concrete situations that need a human.
* **Completion guidance:** how to check that the customer has what they need.

Keep instructions applicable across the agent's enabled apps. Use [skills](/product/ai-agents/skills.md) for detailed procedures that only apply to particular requests.

## Example: login troubleshooting

```text theme={null}
You specialize in helping customers troubleshoot login problems.

Use a calm, friendly tone. Give one clear next step at a time and ask
for the exact error message and platform when that information is missing.
Use the published knowledge base for product-specific troubleshooting.

Help with documented sign-in and connection checks. Do not ask for
passwords or authentication codes, and do not claim to change an account.

Escalate account recovery, ownership disputes, suspected account theft,
and requests that need an account change. If the customer asks for a human,
hand the ticket over.

Do not repeat a troubleshooting step the customer already tried without
explaining why it is useful to try it again. After offering help, check
whether it worked. Keep helping with follow-up questions until the issue
is handled or a human is needed.
```

Adapt this example to your actual support policy and published troubleshooting guidance.

## How the agent helps

You do not need to repeat the agent's usual behavior in your instructions. It already aims to:

* Answer from your published knowledge base in the customer's language.
* Ask follow-up questions and keep helping while the ticket remains open.
* Hand over requests that need human support, such as account changes, compensation, or a request to speak to a person.
* Close the ticket when the issue is handled and the customer confirms they need no further help.

Use your instructions to add specific support policies and handoff conditions.

## Make guidance actionable

Prefer concrete conditions over vague instructions such as “escalate difficult cases.” For example:

```text theme={null}
Escalate when restoring progress requires an account change, or when
the customer reports missing items that need a purchase investigation.
```

Explain the desired behavior as well as the boundary:

```text theme={null}
When escalating a missing purchase, explain that a human needs to review
it. Do not promise a refund, replacement, or completion time.
```

For requests that need account changes or refunds, describe when to hand over to a human. Instructions cannot enable the agent to perform those actions itself.


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.