Frustrated developer lets AI improve its own code, yielding a Word document converter that trounces existing tools

A developer fed up with sluggish and error-prone backend Word document generation has created dom-docx, an open-source library that converts HTML into native, editable Word documents using an AI-driven self-improvement loop. According to ShortSingh, the project employed 'Autoresearch loops,' a technique in which an agent autonomously iterates and self-improves against an objective scoring metric.

ShortSinghDeveloper builds open-source HTML-to-DOCX converter using AI self-improvement loops

"The scoring system evaluated each conversion across 37 real-world HTML test cases, weighing visual layout fidelity at 50%, document editability at 35%, and compile speed at 15%."

This automated feedback loop allowed the library to rapidly refine its output without manual tuning, a stark contrast to the usual approach of hand-patching conversion rules.

When benchmarked against established npm HTML-to-DOCX libraries, dom-docx achieved average layout-fidelity scores in the mid-90s percent, while competing tools hovered around the mid-60s percent, ShortSingh reported. The result is a 30-percentage-point gap that the developer attributes directly to the AI's iterative optimization.

Hacker News noted that dom-docx converts semantic HTML fragments into native, editable Word documents (OOXML format) and supports paragraphs, runs, lists, tables, and images. The project includes a live demo and a visual regression testing loop that renders HTML in Chromium, converts to docx, rasterizes via LibreOffice, and scores layout and structural fidelity against a human-validated metric. As Hacker News put it, the tool uses "actual document elements rather than screenshots or layout hacks." :: Hacker News

That emphasis on real document structure means users can edit the resulting Word files as normal, rather than wrestling with embedded images or fragile positioning.

Dom-docx requires Node.js 20 or later and optionally uses Playwright for certain conversion paths, according to Hacker News. The library is available on GitHub, along with its scoring framework, making it possible for other developers to adapt the AI self-improvement approach to their own document-generation challenges.