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

# Prompt Management

> Store, version, and serve prompts dynamically from the dashboard.

Hardcoding complex system prompts in your application code makes iteration slow. Every prompt tweak requires a code review, a build, and a deployment.

MetrixLLM's **Prompt Registry** solves this by separating prompt engineering from your codebase. You manage the text of your prompts in the dashboard, and your code simply references them by ID.

## Creating a prompt

In the dashboard, navigate to **Prompts** and click **New Prompt**.

Write your system prompt, including variables wrapped in `{{ }}` brackets. For example:

> `You are a helpful customer support agent for {{ company_name }}. Answer the following question: {{ user_query }}`

When you save, the prompt is assigned a unique UUID (the `prompt_id`), and version `v1` is created.

## Using a prompt in your code

To use the prompt, you send a standard request to the gateway, but instead of passing the raw messages array, you pass the `prompt_id` and the `inputs` object.

The gateway resolves the prompt template, injects the variables, formats the messages, and forwards it to the provider.

<CodeGroup>
  ```json OpenAI format theme={null}
  {
    "model": "gpt-4o",
    "prompt_id": "123e4567-e89b-12d3-a456-426614174000",
    "inputs": {
      "company_name": "MetrixLLM",
      "user_query": "How do I reset my password?"
    }
  }
  ```

  ```json Anthropic format theme={null}
  {
    "model": "claude-opus-4-5",
    "max_tokens": 1024,
    "prompt_id": "123e4567-e89b-12d3-a456-426614174000",
    "inputs": {
      "company_name": "MetrixLLM",
      "user_query": "How do I reset my password?"
    }
  }
  ```
</CodeGroup>

## Versioning

Every time you edit a prompt in the dashboard, a new version is created (`v2`, `v3`, etc).

The gateway always resolves requests to the **latest version** of the prompt. If you push a bad prompt, you can instantly revert live traffic by going to the prompt's version history and clicking **Set as Latest** on an older version.

<Note>
  For performance, the gateway caches resolved prompt templates in memory. Changes made in the dashboard may take up to **5 minutes (300s)** to fully propagate to all live traffic.
</Note>