It reminds me of this paper from MIRI a few years ago discussing models which treat themselves as an explicit part of the environment. I think it's a very productive approach - https://arxiv.org/abs/1902.09469
This is wild projection, but it reminds me of the good advice I got when writing undergrad philosophy papers – map out your tree of arguments (especially if they're logically complicated), and structure your writing around that; talk in as conversational prose as you can because it mercilessly exposes jargon; highlight when you introduce new concepts.