> ## Documentation Index
> Fetch the complete documentation index at: https://docs.0mcp.io/llms.txt
> Use this file to discover all available pages before exploring further.

# Prompts

> Learn how MCP prompts provide reusable, user-controlled message templates with arguments and optional embedded context.

**[MCP prompts](https://modelcontextprotocol.io/specification)** are reusable message templates that an MCP server exposes to a client.

They help users start common AI workflows with consistent instructions and context. A prompt might prepare a code review, summarize an incident, analyze a customer account, or draft a project update.

In MCP's control model, prompts are **user-controlled**. The user explicitly chooses a prompt, often through a slash command, menu item, or command palette.

## How MCP prompts work

A typical prompt flow is:

1. The client discovers prompts with `prompts/list`.
2. The user selects a prompt.
3. The client requests it with `prompts/get` and supplies any arguments.
4. The server returns one or more structured messages.
5. The client adds the messages to the AI interaction.

```text theme={null}
User         Client                         Server
 |             |---- prompts/list ----------->|
 |             |<--- available prompts --------|
 |-- selects ->|---- prompts/get + arguments ->|
 |             |<--- structured messages -------|
```

Servers declare the `prompts` capability during initialization when they support this feature.

## What defines a prompt?

A prompt definition can include:

* A unique programmatic `name`
* An optional display `title`
* A human-readable `description`
* Optional arguments
* Optional icons and metadata supported by the protocol version

```json theme={null}
{
  "name": "review_pull_request",
  "title": "Review pull request",
  "description": "Review a pull request for correctness, security, and maintainability.",
  "arguments": [
    {
      "name": "pull_request_url",
      "description": "The full URL of the pull request.",
      "required": true
    }
  ]
}
```

Arguments let one prompt template adapt to a specific repository, customer, date range, language, or other user-selected input.

## What does `prompts/get` return?

The server returns a description and an ordered list of prompt messages. Each message has a `user` or `assistant` role and supported content.

```json theme={null}
{
  "description": "Review pull request 1042",
  "messages": [
    {
      "role": "user",
      "content": {
        "type": "text",
        "text": "Review this pull request. Identify correctness, security, and maintainability issues."
      }
    }
  ]
}
```

Prompt messages can contain text and other content types supported by the negotiated protocol version. They can also embed server-provided resources so the workflow arrives with relevant context.

## Prompts, resources, and tools

| MCP primitive | Main purpose                 | Typical controller |
| ------------- | ---------------------------- | ------------------ |
| **Prompt**    | Start a reusable AI workflow | User               |
| **Resource**  | Supply readable context      | Application        |
| **Tool**      | Perform an operation         | Model              |

A prompt can guide the model to use [resources](/learn/core-concepts/resources) or [tools](/learn/core-concepts/tools), but it does not replace them.

For example, a “Prepare account review” prompt can define the workflow. A customer resource can supply account context. A tool can retrieve current usage or create a follow-up task.

## Examples of MCP prompts

| Prompt                   | Arguments               | Intended result              |
| ------------------------ | ----------------------- | ---------------------------- |
| `review_pull_request`    | Pull request URL        | Structured code review       |
| `summarize_incident`     | Incident ID             | Timeline and lessons learned |
| `prepare_customer_brief` | Customer ID             | Account briefing             |
| `explain_error`          | Error message, language | Plain-English diagnosis      |
| `draft_release_notes`    | Version, audience       | Release-note draft           |

## Design prompts for clear selection

* Use an action-oriented name and title.
* Explain when the user should choose the prompt.
* State the expected output.
* Keep each argument focused and well described.
* Validate required arguments before generating messages.
* Keep instructions concise and specific.
* Avoid silently adding unrelated goals.
* Test prompts with missing, empty, and unusual inputs.

Prompt output should be treated as untrusted content. A prompt can influence model behavior, so clients and servers should avoid inserting secrets and should make the generated messages visible when appropriate.

## Prompt updates

A server can declare `listChanged` support. If the available prompt list changes, it can send `notifications/prompts/list_changed` so the client can refresh its view.

Clients should not assume every server or protocol version supports the same metadata or content types. They should follow negotiated capabilities.

## Key takeaway

**An MCP prompt is a user-selected, reusable template that returns structured messages for a consistent AI workflow.**