Copy from ChatGPT, paste into the converter, click DOCX. Headings, lists, code blocks, tables, and LaTeX equations all carry over. Same flow works for Claude, Gemini, and Perplexity.
pandoc -f gfm+tex_math_dollars+footnotes -t docx --mathml), with a fully-local docx-package fallback.$x$) and block ($$x$$) translate to native Office Math (OMML), fully editable in Word./api/convert-to-docx; temp files deleted after each request.ChatGPT, Claude, and Gemini all expose a small copy icon next to each response. Use it. Manual text selection silently strips markdown — code fences, list bullets, and LaTeX delimiters all vanish.
If you've already lost formatting to a manual selection, regenerate the response and copy with the button. There's no recovery path for stripped markdown.
Visit the homepage. No signup, no install — the converter loads as a single page and runs entirely in your browser.
No signup or install — the converter runs entirely in your browser.
Cmd-V (Mac) or Ctrl-V (Windows/Linux). The right panel renders a paper-clean preview as you type — headings, lists, tables, syntax-highlighted code, and KaTeX equations all live.
The preview is what your DOCX/PDF will look like. If something's wrong here, fixing it in the markdown is faster than fixing it in Word.
Hit DOCX in the export bar. The file downloads instantly with a timestamped filename. Open in Word, Google Docs, Pages, or LibreOffice — all read it fluently.
Behind the scenes, the markdown POSTs to /api/convert-to-docx (server-side Pandoc) for highest fidelity — better tables, cleaner equations. Pandoc reads it from a /tmp file, produces the DOCX, the temp files get deleted, the binary streams back. If the server route is unreachable, the converter falls back to a fully-local DOCX path using the `docx` package. Both are solid.
Citations / footnotes
Perplexity and Claude sometimes emit numbered citations as Unicode superscripts or markdown footnotes. Footnote-style markdown (`[^1]`) round-trips cleanly; superscripts are kept as text but lose the link.
Image attachments
Inline images render in the export only if the source URL serves with permissive CORS. Generated DALL-E images via OpenAI's CDN do; private workspace uploads typically don't.
Tool-call traces
Function/tool blocks from agent-style chats render as code blocks. They look fine but lose the structural distinction from regular code.
When ChatGPT emits LaTeX ($E = mc^2$ or$$ \int_0^\infty e^{-x^2} dx $$), the DOCX exporter translates each equation to native Office Math (OMML). Word treats them as real equation objects — you can click in and edit any term, change the font, or copy them into another document. PDF and HTML render the same equations via KaTeX.
Both are legitimate. The built-in export is great for simple prose when your mode supports it; conversion wins when fidelity, reliability, or another assistant is involved.
| ChatGPT’s own export | Copy Markdown → convert | |
|---|---|---|
| Availability | Needs the code / data-analysis tool — not in every mode, model, or tier | Any chat, any tier — anything that lets you copy the reply |
| Download link | Sandbox file; the link expires when the session ends | File is generated on your machine at click time — nothing to expire |
| Equations | Often plain text or images in the .docx | Native Office Math (OMML) — click into any equation in Word and edit it |
| Tables | Simple tables usually fine; complex ones frequently break | Real Word tables with alignment, via Pandoc |
| What you get | Generated by the model — you see the result after opening it | The live preview is the document; download matches it exactly |
| Other assistants | ChatGPT only | Same flow for Claude, Gemini, Perplexity, and any Markdown |
Almost every failed ChatGPT→Word export is one of these. Match the symptom, apply the fix.
Files generated by ChatGPT's code tool live in a sandbox that resets when the session ends — old download links die with it (“file not found”). Asking ChatGPT to regenerate works, but the faster path is to skip the sandbox entirely: copy the response as Markdown and convert it locally, so there's no link to expire.
Word pastes the raw text and throws away the Markdown that carried the structure — headings flatten, bullets become hyphens, tables become pipe characters. Don't paste into Word directly: paste into a Markdown converter and open the DOCX it produces. Every heading, list, and table arrives as real Word formatting.
That's the raw Markdown table syntax. Word doesn't parse it, but any Markdown-to-DOCX converter does — the pipes become a real table with aligned columns you can style in Word.
ChatGPT wraps math in \( … \) / \[ … \] or bare parentheses instead of the $ delimiters most tools expect. LLM to Doc's preprocessor detects and rewrites these automatically; elsewhere, search-and-replace them with $ / $$ before converting.
Browser clipboard limits and lazy-loaded chat history can clip very long conversations. Copy response-by-response with each message's copy button instead of selecting the whole thread — the per-message button always captures the complete Markdown.
Use ChatGPT's built-in export when it's available in your mode and the content is simple prose — it's one prompt and you're done. Use a converter when the document has equations or complex tables, when the download link has expired, when you're on a tier or mode without the code tool, or when you want to see exactly what the file will contain before downloading. The converter route also works identically for Claude, Gemini, and Perplexity.
Yes — ChatGPT can generate a .docx with its code / data-analysis tool (and some custom GPTs), giving you a download link. Two caveats: it isn't available in every mode or tier, and the auto-generated formatting is often basic — complex tables and LaTeX equations frequently don't survive. And if you just copy a reply and paste it into Word, the Markdown structure is lost. For a faithful document — editable Office Math equations, real tables, highlighted code — converting the Markdown with a tool like LLM to Doc is more reliable.
Copy the response using ChatGPT's copy button, paste it into LLM to Doc, and click DOCX. It takes about three minutes end-to-end, it's free, and there's no signup or extension to install.
Pasting raw text into Word discards the Markdown that carries the structure — fenced code, bullet markers, and LaTeX delimiters all vanish. Converting the Markdown to DOCX instead of pasting rebuilds that structure as real Word formatting.
Yes. Tables become real Word tables with aligned columns, and LaTeX math ($…$ and $$…$$) becomes native, editable Office Math (OMML) — you can click into any equation in Word and edit it.
LLM to Doc is a free web app — no browser extension, signup, or quota. The same flow also works for Claude, Gemini, Perplexity, and any assistant that emits Markdown.
The usual causes: the .docx download link from ChatGPT's code tool expired with the session (“file not found”), the formatting was stripped because the text was pasted straight into Word, or equations arrived as \( … \) text instead of rendered math. The reliable workaround for all three is the same — copy the response as Markdown with the copy button and convert it to DOCX instead of relying on ChatGPT's own file generation.
Files created by ChatGPT's code / data-analysis tool are stored in a temporary sandbox that resets when your session ends, so download links go dead — often within an hour or two. Regenerating the file works temporarily; converting the Markdown yourself avoids the expiring link entirely.
Paste a ChatGPT response and you’ll have a clean DOCX in under five seconds.
Open the converter