Compile LLM Agent Prompts with MDS CLI

What you’ll get

You will learn how to use a tool called MDS (Markdown Script) to write your AI agent prompts in one place and automatically convert them into different formats needed by various AI platforms. This saves you from copying and pasting prompts, reducing errors and making updates easier.

Tools you need

• mds-cli (free): A command-line tool that compiles your special Markdown prompts into formats AI models understand.
• Rust (free): A programming language used to build the mds-cli tool. You need it to install the tool.
• Makefile (free): A tool that helps automate build processes, often used with Rust projects.

Steps

1. Install the mds-cli tool: You need to install the command-line tool first. This requires having Rust installed on your computer.

   • If you don’t have Rust installed, visit the official Rust website (rust-lang.org) and follow their installation instructions. This is a one-time setup.
   • Once Rust is installed, open your terminal or command prompt and run the following command to install mds-cli:

   cargo install mds-cli

   You should see messages indicating the tool is being downloaded and installed. Once complete, the command prompt will return to a ready state.

2. Create your prompt file: Create a new file for your prompt. The author uses a .mds extension for these special Markdown files. Let’s call it review.mds.

3. Write your prompt using MDS syntax: Open review.mds in a text editor. You can use variables, functions, and imports. Here’s an example based on the author’s description:

   First, define a reusable persona in a separate file, for example, _persona.mds:

   @define reviewer(lang):
     You are a senior {lang} code reviewer.
     Flag correctness bugs first, then style.
     Keep it concise.
   @end

   Then, in your main review.mds file, import and use this persona. You can also define message blocks that will compile to JSON format for AI models:

   @import "./_persona.mds" as p
   
   @message system:
     {p.reviewer("TypeScript")}
   @end
   
   @message user:
     Review this diff:
     {diff}
   @end

   You should see the text you typed appear in your editor. The special @ commands are how MDS understands your prompt structure.

4. Compile your prompt: Go back to your terminal or command prompt, navigate to the directory where you saved your .mds files, and run the build command. The author shows mds build review.mds.

   mds build review.mds

   If your review.mds file contains @message blocks, this command will create a review.json file in the same directory. This JSON file contains the structured messages ready to be sent to an AI model. If there were no @message blocks, it would create review.md.

   You should see a new file, review.json, appear in your file explorer. Opening it will show a list of message objects, like this example:

   [
     {
       "role": "system",
       "content": "You are a senior TypeScript code reviewer. Flag correctness bugs first, then style. Keep it concise."
     },
     {
       "role": "user",
       "content": "Review this diff: \n ..."
     }
   ]

   If your file only contained Markdown elements (no @message blocks), it would create review.md with the compiled Markdown content.

Original source

This workflow is based on a blog post by dean0x, explaining how to use their custom tool, MDS (Markdown Script), to simplify writing and managing prompts for AI agents. The goal is to write prompts once and have them work across different AI platforms.

Notes & variations

• Free tier alternatives: While MDS is a specific tool, the core idea of managing prompts can be approached by using simple text files and carefully organizing them, though it lacks the automation and error checking MDS provides.
• Common mistake: Forgetting to install Rust before trying to install mds-cli will cause the cargo install command to fail with an error message about cargo not being found.
• Tip for better results: Use partial files (files starting with _) for common prompt components like personas or instructions. This allows you to update a single file and have all prompts that import it automatically recompile with the changes.