There are three main approaches to migrate your existing LLM application to AgentMark:

​
1. Using Serialization

The fastest way to migrate is using AgentMark’s serialization utilities to convert your existing completion parameters into AgentMark files:

TypeScript
import { serialize } from "@puzzlet/agentmark";

// Your existing completion params
const params = {
  model: "gpt-4",
  messages: [
    { role: "system", content: "You are a helpful assistant" },
    { role: "user", content: "Hello!" }
  ],
  temperature: 0.7
};

// Convert to AgentMark
const mdx = serialize(params, "my-prompt");

// Save to file
fs.writeFileSync("my-prompt.prompt.mdx", mdx);

​
2. Importing Markdown Files

If you have existing markdown files containing prompts (without frontmatter), you can convert them to AgentMark format by importing them into your prompt file:

example.prompt.mdx
---
name: example
metadata:
  model:
    name: gpt-4
    settings:
      temperature: 0.7
---
import ExistingSystemPrompt from "./existing-system-prompt.md";
import ExistingUserPrompt from "./existing-user-prompt.md";

<System>
  <ExistingSystemPrompt />
</System>
<User>
  <ExistingUserPrompt />
</User>

​
3. Manual Migration

For more control, you can manually migrate your prompts following these steps:

  1. Create a new .prompt.mdx file
  2. Add frontmatter configuration:
---
name: my-prompt
metadata:
  model:
    name: gpt-4
    settings:
      temperature: 0.7
---
  1. Convert messages to AgentMark components:
<System>Your system message here</System>
<User>Your user message here</User>
<Assistant>Your assistant message here</Assistant>
  1. Replace hardcoded values with props:
<User>Hello {props.name}!</User>

​
Example Migration

Before (OpenAI format):

const completion = await openai.chat.completions.create({
  model: "gpt-4",
  messages: [
    { role: "system", content: "You are a helpful assistant" },
    { role: "user", content: `Hello ${username}!` }
  ],
  temperature: 0.7
});

After (AgentMark):

---
name: greeting
metadata:
  model:
    name: gpt-4
    settings:
      temperature: 0.7
---

<System>You are a helpful assistant</System>
<User>Hello {props.username}!</User>

import { runInference, load } from "@puzzlet/agentmark";

const prompt = await load("./greeting.prompt.mdx");
const result = await runInference(prompt, { username: "Alice" });

​
Best Practices

  1. Start with a small subset of prompts to test the migration
  2. Use serialization when possible to quickly convert existing prompts
  3. Manually migrate complex prompts with custom logic
  4. Make use of reusable components to keep your prompts clean and organized
  5. Test prompts after migration to ensure behavior matches

Have Questions?

We’re here to help! Choose the best way to reach us:

Was this page helpful?

Model ProvidersArchitecture