I do a lot of backend document (docx file) generation work, and updating our templates and backend code is one of my least favorite development tasks. Cryptic errors, minute-plus rebuild loops. I’d much prefer building these reports in JS rendered HTML (e.g., Vue or React), but existing HTML-to-docx libraries in the OSS ecosystem don't produce output that's actually valid, editable Word structure.
I'd had good luck applying Karpathy's Autoresearch pattern (agent runs iterations against an objective score, keeps what improves, discards what doesn't) to a couple of other problems, and figured OOXML fidelity was a good fit.
The Autoresearch process goes like this: render HTML and take a screenshot, use dom-docx to convert to docx, rasterize the docx file with LibreOffice and take another screenshot, score the browser HTML screenshot vs LibreOffice screenshot and measure layout fidelity + editability + speed as a quality metric, feed score back in, and repeat to drive higher fidelity within the constraints of editability and performance. It’s a beautiful thing to watch.
Burned some tokens and ran that loop against 37 real-world HTML patterns such as nested lists, tables, flex layouts and blockquotes (stuff that often breaks converters) and "brute forced" my way to what I hope is a high-fidelity HTML to docx converter.
A few things about where it landed:
- Native OOXML output, real Word structure, not a screenshot or a 1x1 table pretending to be a document
- Works in Node, in the browser (no Playwright needed for the default path), and as a CLI (npx dom-docx input.html -o output.docx)
- MIT licensed
- Full benchmark methodology + results vs the established OSS alternatives: https://github.com/floodtide/dom-docx/blob/main/docs/BENCHMA...
Live demo if you want to test HTML conversion in the browser: https://dom-docx.com
Happy to answer anything about the scoring loop or anything else.
PS: This is the first thing I've open sourced and I'm excited to see where it leads!
Sounds to good to be true, but it works for the given examples!
What is the scope though? It produces garbage with larger documents with tables and figures. For example this large document, even with removed menu and headings: https://docs.redhat.com/en/documentation/red_hat_enterprise_...
All of the suite test cases are small fragments of HTML to test specific things, so I probably do need a way to test bigger documents. Thanks for the feedback. I’ll add that to the to-do list.
That's an interesting approach. I'm concerned about the use of LibreOffice as your source of truth. Would it be possible to swap out LibreOffice for actual MS Word in this workflow? This also could reveal some libreoffice rendering bugs/edge cases that are worth filing bugs over.
I don’t have MS word installed on my development machine, so I haven’t done much testing with Word, but that does sound like a good idea to run the suite using Word. For what it’s worth I did not notice any issues with LibreOffice, even after running many iterations and tests.
I would add a +1 for testing w/ Word - the official Office suite runs some validation where only Word will show a "broken file" popup, even when nothing else does.
In our case, clients use only real Word, so any machine-generated/mutated files (excel/ppt as well) need a pass through the real office executable.
Interesting but how is an "autoresearch loop" different than creating a spec and X number of testcases and letting an agent run against these testcases and the spec?
Just adding, since I don't see it mentioned in the readme, that its written in Typescript. That's what makes this interesting. I imagine Pandoc can do this but its not Typescript (its Haskell).
Thanks, and just to mention, I’ve had awesome results using Autoresearch loops on other things like SQL performance.
Prompt example:
You are a SQL performance researcher. Run the following SQL query to establish a baseline, then come up with hypotheses to improve performance. Score each result and run 5 iterations. Avoid any regressions, each result must contain the exact same rows and columns.
Also works wonderful for generating AI scripts, goal being increasing the elo rating after a tournament run. Also having deep and shallow tests save a lot of time. Deep tests are run sparingly while shallow tests are run after each change.
I have done something similar for PPTX, keeping the fidility intact was really challenging with computed values & OOXML counterparts and again challanges with different XML implementations like that of MS & Libre..
I did everything with the Cursor $20 plan and the Claude Code $20 plan, so in this particular case it wasn’t actually that many tokens. This was over the course of about four weeks of weekend and evening work.
Hey HN, author here.
I do a lot of backend document (docx file) generation work, and updating our templates and backend code is one of my least favorite development tasks. Cryptic errors, minute-plus rebuild loops. I’d much prefer building these reports in JS rendered HTML (e.g., Vue or React), but existing HTML-to-docx libraries in the OSS ecosystem don't produce output that's actually valid, editable Word structure.
I'd had good luck applying Karpathy's Autoresearch pattern (agent runs iterations against an objective score, keeps what improves, discards what doesn't) to a couple of other problems, and figured OOXML fidelity was a good fit.
The Autoresearch process goes like this: render HTML and take a screenshot, use dom-docx to convert to docx, rasterize the docx file with LibreOffice and take another screenshot, score the browser HTML screenshot vs LibreOffice screenshot and measure layout fidelity + editability + speed as a quality metric, feed score back in, and repeat to drive higher fidelity within the constraints of editability and performance. It’s a beautiful thing to watch.
Burned some tokens and ran that loop against 37 real-world HTML patterns such as nested lists, tables, flex layouts and blockquotes (stuff that often breaks converters) and "brute forced" my way to what I hope is a high-fidelity HTML to docx converter.
A few things about where it landed:
- Native OOXML output, real Word structure, not a screenshot or a 1x1 table pretending to be a document - Works in Node, in the browser (no Playwright needed for the default path), and as a CLI (npx dom-docx input.html -o output.docx) - MIT licensed - Full benchmark methodology + results vs the established OSS alternatives: https://github.com/floodtide/dom-docx/blob/main/docs/BENCHMA...
Live demo if you want to test HTML conversion in the browser: https://dom-docx.com
Happy to answer anything about the scoring loop or anything else.
PS: This is the first thing I've open sourced and I'm excited to see where it leads!
Sounds to good to be true, but it works for the given examples! What is the scope though? It produces garbage with larger documents with tables and figures. For example this large document, even with removed menu and headings: https://docs.redhat.com/en/documentation/red_hat_enterprise_...
Could it be it has only been fed small documents?
All of the suite test cases are small fragments of HTML to test specific things, so I probably do need a way to test bigger documents. Thanks for the feedback. I’ll add that to the to-do list.
That's an interesting approach. I'm concerned about the use of LibreOffice as your source of truth. Would it be possible to swap out LibreOffice for actual MS Word in this workflow? This also could reveal some libreoffice rendering bugs/edge cases that are worth filing bugs over.
I don’t have MS word installed on my development machine, so I haven’t done much testing with Word, but that does sound like a good idea to run the suite using Word. For what it’s worth I did not notice any issues with LibreOffice, even after running many iterations and tests.
I would add a +1 for testing w/ Word - the official Office suite runs some validation where only Word will show a "broken file" popup, even when nothing else does.
In our case, clients use only real Word, so any machine-generated/mutated files (excel/ppt as well) need a pass through the real office executable.
Thanks for the feedback, I’ll add Word validation to my todo list.
Interesting but how is an "autoresearch loop" different than creating a spec and X number of testcases and letting an agent run against these testcases and the spec?
Just adding, since I don't see it mentioned in the readme, that its written in Typescript. That's what makes this interesting. I imagine Pandoc can do this but its not Typescript (its Haskell).
The screenshot-to-docx scoring loop is a clever way to verify layout fidelity. Very useful for anyone generating reports from HTML.
Thanks, and just to mention, I’ve had awesome results using Autoresearch loops on other things like SQL performance.
Prompt example:
You are a SQL performance researcher. Run the following SQL query to establish a baseline, then come up with hypotheses to improve performance. Score each result and run 5 iterations. Avoid any regressions, each result must contain the exact same rows and columns.
See https://github.com/karpathy/autoresearch
Also works wonderful for generating AI scripts, goal being increasing the elo rating after a tournament run. Also having deep and shallow tests save a lot of time. Deep tests are run sparingly while shallow tests are run after each change.
I have done something similar for PPTX, keeping the fidility intact was really challenging with computed values & OOXML counterparts and again challanges with different XML implementations like that of MS & Libre..
Agreed, visual fidelity is pretty hard, and that’s why none of the scores are 100.
this is cool. I'm curious how many tokens ($) it cost you to build?
i'm also building in this space (an MCP for agents to manipulate docx)
I did everything with the Cursor $20 plan and the Claude Code $20 plan, so in this particular case it wasn’t actually that many tokens. This was over the course of about four weeks of weekend and evening work.
Good work on the effort to the author.
And every time y'all, how absurd is it that this still has to exist? It never stops being wild to me.
Surely I can't be the only one who misread the heading and clicked expecting to see DOOM implemented using just HTML DOM. :-/
Pretty sure that has been done. I clicked to see DOOM implemented using OOXML.
For that you probably want this, lol: https://cssdoom.wtf/