I currently have subscriptions to both Claude and ChatGPT. I generally prefer the former but find I can't fully commit to it for my maths-heavy workload as it often struggles to correctly render LaTeX.
An example of this failed rendering is here [1]. If I use Claude for all of my work, I come across issues like this or worse at least once a day. Instead, I find it easier just to ask any maths questions to ChatGPT which seems to have a much more robust system for outputting LaTeX.
I would love to merge my subscriptions though, so I'm here to ask whether anyone has a system prompt that has been effective in guiding Claude towards producing valid LaTeX. I've tried a few prompts myself but struggled to find anything that it consistently followed.
[1] https://imgur.com/yzlluOA
ontouchstart3 hours ago
LLMs only produce markdown [1], usually Math is wrapped with KaTeX [2]. The rendering happens in the Web UI. If some math failed to render, you can copy the code and paste to latex-sandbox [3] and fix yourself.
[1] https://docs.github.com/en/get-started/writing-on-github/wor...
Johnny_Bonk3 hours ago
When you say consistently render what do you mean? Like without any sort of formatting issue? If thats the case I would lower your expectations a bit almost always as with any bit of copy editing regardless of using ai or not theres going to be issues to review. I have a claude skill using tectonic I believe and it works pretty well, and then I wrote hooks to make sure if certain things are going to be written it does it in a certain way.
ontouchstart3 hours ago
It seems that GitHub gist can renders some of \LaTeX, but not perfect.
https://gist.github.com/ontouchstart/bcffb186a753c5b75522fc8...
jlongo782 hours ago
honestly the magic phrase ive found is "use aligned LaTeX delimiters and assume the renderer supports MathJax." that one line cuts 80% of the inconsistency.
also: put it in your CLAUDE.md, not the system prompt. Claude reads that per-session and it sticks way better than hoping your prompt survives context compression. learned that the hard way after losing formmating mid-session one too many times.
verdverm4 hours ago
Some things are just too hard for them to do reliably