LLM prompts, but structured like a web page

What it does

POML is a markup language for writing complex LLM prompts with semantic tags like <role>, <task>, and <example>. It lets you embed external data—text files, spreadsheets, images—via tags such as <document> and <img>, and includes a built-in templating engine with variables and conditionals. A VS Code extension and Node.js and Python SDKs handle editing, testing, and rendering into application workflows.

The interesting bit

The clever angle is a CSS-like styling system that decouples what a prompt says from how it is formatted. You can tweak verbosity or output syntax through <stylesheet> definitions without rewriting the underlying logic, which directly targets LLM format sensitivity. It treats prompt engineering like front-end development, complete with markup, assets, and style sheets.

Key highlights

• Semantic, HTML-like syntax for prompt components (<role>, <task>, etc.)
• Native embedding of external data types—images, tables, documents
• CSS-inspired styling layer separates content from presentation format
• Built-in templating with variables, loops, and conditionals
• VS Code extension with live preview, diagnostics, and interactive LLM testing
• SDKs available for JavaScript/TypeScript and Python, plus community renderers for Rust, Ruby, and Julia

Caveats

• The VS Code extension’s interactive testing requires manual LLM provider and API key configuration before it functions.

Verdict

Worth a look if you are building data-heavy or multi-modal prompt pipelines and are tired of concatenating strings. Probably overkill if your prompts are still just a few lines of system instructions.