In my very limited experience it helps to tell the model that all commits (both patch and commit message) MUST pass devtools/checkpatches.sh, ideally configured (and verified) as a git hook, clear justification must be provided in rare exception cases.
From the general programming principles I'm unsure about duplicating information already available in doc/guides/contributing in the same tree. Perhaps we could just instruct the model to read patches.rst and coding_style.rst, they are not very big (and models typically know Linux and PIP8 rules already). Otherwise we risk one of the documents becoming outdated, or rules/nuances omitted in the summarized version ignored. In instructions specifically for the model we could focus on things that are persistent source of problems for it despite instructions for humans being clear. Just my 2c, great that someone is working on this.
