What is POML?

• ► Is Microsoft’s new XML/HTML-like language designed to bring structure and clarity to prompt engineering for Large Language Models (LLMs).
• ► It organizes prompts into modular, reusable components, making them easier to maintain and adapt at scale

What’s Novel About It?

• ► Structured Syntax: Semantic tags like role, task, and example allow developers to break down complex prompts into well-defined sections.
• ► Data embedding: Direct support for including documents, tables, and images (document, table, img), enabling richer multimodal prompts.
• ► Separation of Content and Style: CSS-like styling lets you adjust tone, verbosity, or formatting without rewriting the core prompt.
• ► Templating Logic: Built-in support for variables (let), loops (for), and conditionals (if) provides dynamic prompt generation.
• ► Tooling Ecosystem: VS Code Extension: syntax highlighting, auto-completion, previews, and diagnostics. SDKs for Node.js and Python: easy integration into real applications.

A Structured Language for Prompt Engineering

While XML and JSON are widely used for representing structured data in a general sense, they lack a standardized way to capture the unique needs of prompt engineering. POML builds on the familiar XML style but introduces prompt-specific semantics, templating logic, styling separation, and dedicated tooling. The table below illustrates how POML differs from traditional formats by focusing specifically on the requirements of LLM workflows.

Why this is different from plain xml or json?

POML helps teams manage large prompt libraries with clearer structure, reuse, and multimodal support, while tooling like IDE extensions and SDKs speeds up development. Many see it as a step toward treating prompt engineering more like software engineering, even if models themselves don’t inherently “understand” it—the real benefit lies in helping developers stay organized and scale their work. More than just markup, POML represents an effort to standardize and professionalize prompt design, and in a modest way, it may also shape the future of education by making AI interactions more structured, transparent, and accessible to learners and teachers alike

XML vs JSON vs POML

XML and JSON can already represent structured data and even prompts, but POML standardizes a schema and provides dedicated tooling aimed specifically at Large Language Model workflows.

Example: POML in Action

Below is a simple .poml (basically an xml) file describing a prompt from my domain about Parkinson’s Disease, and a Python script that loads it, renders it into a prompt string, and sends it to OpenAI’s API.

        
          role: You are a medical expert explaining Parkinson’s Disease.
          task: Explain the main symptoms of Parkinson’s Disease to a first-year medical student.
          output-format: Use clear language. Provide 4 bullet points. End with one short sentence about why early diagnosis matters.
        
            

run_poml.py

import os
import xml.etree.ElementTree as ET
from openai import OpenAI

# Path to your POML (XML) file
POML_PATH = "/home/ilab/test.pml/pd_poml.xml"

# --- Parse POML (XML) ---
root = ET.parse(POML_PATH).getroot()

def text(tag):
    el = root.find(tag)
    return (el.text if el is not None and el.text else "").strip()

role = text("role")
task = text("task")
outfmt = text("output-format")

# Build prompt string
parts = []
if role:   parts.append(role)
if task:   parts.append(f"\nTASK:\n{task}")
if outfmt: parts.append(f"\nOUTPUT FORMAT:\n{outfmt}")
prompt = "\n".join(parts).strip()

# --- OpenAI API ---
api_key = os.getenv("OPENAI_API_KEY", "").strip()
if not api_key:
    raise RuntimeError("OPENAI_API_KEY is not set. Run: export OPENAI_API_KEY='sk-...'")
if any(ord(c) > 127 for c in api_key):
    raise ValueError("OPENAI_API_KEY is not valid")

client = OpenAI(api_key=api_key)

resp = client.chat.completions.create(
    model="gpt-4o-mini",
    messages=[{"role": "user", "content": prompt}],
)

print("=== POML Prompt Output ===")
print(resp.choices[0].message.content)
    

  PD gait assessment
  postgraduate
  rehabilitation student

  You are a faculty tutor guiding a {{level}} {{audience}}.
  
    Explain how Parkinson’s Disease affects gait and outline a basic assessment protocol.
    Include red flags that warrant referral and 2 short case-study prompts for practice.
    Emphasize {{topic}} and clinical reasoning.
  
  
    - Short overview (3–4 sentences)
    - Bullet list: assessment steps (5–6 items)
    - Red flags (3 items)
    - Two case-study prompts (1–2 sentences each)