Re: [MAINTAINERS SUMMIT] The place of AI code review in the Linux Kernel process
Roman Gushchin <[email protected]> Fri, 17 Jul 2026 14:14:49 -0700
| Newsgroups | dev.linux.lists.ksummit |
|---|---|
| Message-ID | <[email protected]> |
> On Jul 17, 2026, at 2:05=E2=80=AFPM, Krzysztof Kozlowski <[email protected]>= wrote: >=20 > =EF=BB=BFOn 15/07/2026 18:55, Roman Gushchin wrote: >> * Prompt development and testing. >> Currently, prompts are maintained in two GitHub repositories and are >> changed manually or with the help of AI coding agents. However, there is >> no established practice for testing them, especially across various LLM >> models. At the last LSFMMBPF conference, there was a discussion about >> moving them into the kernel tree. I see some pros and cons to this >> approach, but the ownership and testing models are not entirely clear. >=20 > Prompts in a meaning how to parse kernel, work with it or produce > reports: yeah, sure. >=20 > Prompts teaching kernel code: I don't think so. I find it really > incorrect approach to create a dedicated and duplicated basic knowledge > for LLMs, instead of focusing on knowledge for developers. I understand > that LLM has to be taught basics of C [1], but we should not duplicate > kernel coding style or kernel rules into LLM-only instructions. Instead: I agree. I also experiment with Sashiko auto-creating prompts based on human feedback in emails and initial results look promising. So maybe we don=E2=80= =99t even need to write too much manually for LLMs. Thanks! >=20 > 1. Fix docs or kerneldoc to have these properly documented, > 2. Instruct LLM to read these documents. >=20 > We do have tribal knowledge and it should be our kernel docs for the > developers, first of all. If the knowledge is so important for LLM that > it must be duplicated in prompts (just read some of the subsystem > specifics at [2]), then maybe better add it to some guide "writing HID > driver" or "common mistakes in Bluetooth", so developer writing such > driver that will find it. And LLM too. >=20 > I keep growing such codified and condensed (keep as short as possible) > rules for subsystem I am part of [3] and LLM should read that one. LLM > does not need dedicated doc, different than developers need. Well, > except teaching C. [1] :) This is awesome actually!=