I have done that in the past (for my master's thesis), there is only one case where you should do that: the documentation of Isabelle and the documentation of Isabelle developments (like the AFP).
There are some people who do that (Makarius Wenzel, e.g., https://sketis.net/2019/11), similarly the "Concrete Semantics". However, this is not a great solution.
The reasons not to do that:
The compilation takes much longer than using pdflatex, even if you base your work on an image of your development.
Unless you type LaTeX macros directly, you are much more limited in what you can do (LaTeX wise). And if you do type latex macros directly, you cannot produce HMTL output anymore. So the gain of Isabelle is limited.
Many conferences want to see LaTeX sources and they don't run Isabelle, so you will have to generate LaTeX at some point anyhow (and even possibly do some post-productions effects because Isabelle is not able to do some things).
You rarely want to use the exact theorem of Isabelle (LaTeXsugar can help, but it is not perfect).
What if you write the paper now and discover a typo you want to fix in 5 years? In 5 years, latex will still work. Isabelle2020 probably not.
Do all your co-authors use Isabelle on all their computers, including your laptop if you are on vacation and have an emergency fix to do?
And you will fight Isabelle a lot, for example:
text "
\begin{counterexample}
"
lemma True
by auto
text "
\end{counterexample}
"
does not work, because text is its own environment, so you need post-production effects.
Basically, use the snippets mechanism to extract LaTeX out of the theories and then use your favorite LaTeX editor.
