I've been building AI systems for a while and kept hitting the same wall - prompt engineering felt like string concatenation hell. Every complex prompt became a maintenance nightmare of f-strings and template literals.

So I built LLML - think of it as React for prompts. Just as React is data => UI, LLML is data => prompt.

The Problem:

  # We've all written this...
  prompt = f"Role: {role}\n"  
  prompt += f"Context: {json.dumps(context)}\n"  
  for i, rule in enumerate(rules):  
      prompt += f"{i+1}. {rule}\n"  
  
  # The Solution:  
  from zenbase_llml import llml  
  
  # Compose prompts by composing data
  context = get_user_context()
  prompt = llml({  
      "role": "Senior Engineer",  
      "context": context,
      "rules": ["Never skip tests", "Always review deps"],
      "task": "Deploy the service safely"
  })

  # Output:  
  <role>Senior Engineer</role>  
  <context>  
    ...  
  </context>  
  <rules>  
    <rules-1>Never skip tests</rules-1>  
    <rules-2>Always review deps</rules-2>  
  </rules>  
  <task>Deploy the service safely</task>  

Why XML-like? We found LLMs parse structured formats with clear boundaries (<tag>content</tag>) more reliably than JSON or YAML. The numbered lists (<rules-1>, <rules-2>) prevent ordering confusion.

Available in Python and TypeScript:

  pip/poetry/uv/rye install zenbase-llml
  npm/pnpm/yarn/bun install @zenbase/llml

Experimental Rust and Go implementations also available for the adventurous :)

Key features:

  - ≤1 dependencies
  - Extensible formatter system (create custom formatters for your domain objects)
  - 100% test coverage (TypeScript), 92% (Python)
  - Identical output across all language implementations

The formatter system is particularly neat - you can override how any data type is serialized, making it easy to handle domain-specific objects or sensitive data.

GitHub: https://github.com/zenbase-ai/llml

Would love to hear if others have faced similar prompt engineering challenges and how you've solved them!