Most markdown-to-Word tools are a wrapper around a library. AI2Word runs a purpose-built pipeline. Here’s exactly what happens between the moment you click Download Word and the moment Word opens your file.
Each stage exists for a reason. Here’s the reason.
UTF-8 BOMs, mixed line endings, zero-width characters, Unicode spaces, and smart quotes are all normalised before anything else touches the text.
Why: a paste from ChatGPT’s web UI carries invisible characters that break every downstream parser. Cleaning once, at the top, saves ten patches further down.
Every fenced block and inline `span` is lifted out
and replaced with an invisible sentinel character.
Why: $x inside a code sample is
not a variable. A naïve math regex doesn’t know that. Shielding
first means $this->db->query() survives
intact.
$…$, $$…$$, \(…\),
\[…\], and \begin{env}…\end{env}
are located. A looksLikeMath guard rejects dollar
amounts, English words, and $ x $ with stray
spaces.
Why: turning every
$5 and tax is $ into an equation is the fastest
way to lose a user’s trust. False negatives are cheap; false
positives are not.
Headings, lists, tables, blockquotes, horizontal rules, and paragraphs are split into a block list. A special case folds ChatGPT’s “N.” on its own line into a numbered list item.
Why: line-based block parsing is deterministic and debuggable. A regex-first approach tangles when tables and lists nest.
Blocks are converted to an intermediate
DocumentNode tree — headings, paragraphs, lists,
tables, equations. Inline spans (bold, italic, links, math)
are parsed within each block.
Why: separating “what the document says” from “how Word writes it” means the same IR could render to DOCX, HTML, or PDF with no parser changes.
A recursive-descent LaTeX parser turns
\frac{a}{b}, \sqrt[n]{x},
\sum_{i=1}^n,
\begin{pmatrix}…\end{pmatrix} into real Office
Math nodes.
Why: Word speaks OMML, not LaTeX. Rendering equations as Unicode approximations loses subscripts, fractions, and matrices. Native OMML means the equation is editable in Word’s equation editor.
Styles, numbering, headers, footers, settings, and content are
written as separate XML parts and zipped into the
.docx container.
Why: a .docx is a ZIP with a
specific internal structure. Writing the parts directly is
faster and more controllable than any library.
The generated file is stored, handed to the browser, and pruned after two hours. Old documents are removed on every download.
Why: conversion happens on the server, but storage doesn’t have to be permanent. Old documents are removed on every download so the disk never grows without bound.
Paste something with a heading, a table, and a formula. Watch the preview.
Open the app