Re: [MAINTAINERS / KERNEL SUMMIT] AI patch review tools

[Chris Mason <clm@meta.com>]

On 10/8/25 5:08 PM, Kees Cook wrote:
> On Wed, Oct 08, 2025 at 02:04:26PM -0400, Chris Mason wrote:
>> Claude really needs aggressive use of checklists to get it to stay
>> focused, other AIs may need other tricks.
> 
> So very very true. "I did the high priority stuff; I'm finished now".
> 
> Your prompts are very detailed! Have you been able to use a feedback loop
> at all to have it update its own prompts to help you clarify its results?

Yeah, I started with a very basic prompt and ran it through ~300 Fixes:
tagged patches with known bugs.  So I had the buggy patch and the fix,
and I'd have AI review the bad patch and then ask it to compare the
review against the known regression.

If the review failed, I'd ask it to fix the prompt in a generic way to
find the bug.  The prompts would grow and grow and then I'd ask AI to
streamline them.  I did those iterations a few times and fixed things up
as it went.

The false positive detection was more manual, that was mostly me asking
claude which part of the prompt made it flag something, and then either
fixing that part of the prompt or adding onto the false positive guide.

> I've found that pretty helpful for development-oriented prompting. i.e.
> "Compare the foo.md file to what you now know about the foo process
> and update it to reflect anything that needs to be added, removed,
> or changed."
> 
> Additionally, have you been able to build any patterns from existing
> bugs? The "Pattern ID" stuff feels like you built this from specific
> instances?
> 
> I'm really curious if the "CONTEXT MANAGEMENT" section actually works? I
> never considered tell it to discard specific context like you have,
> "Discard non-essential details after each phase to manage token limits"
> I'm going to try that immediately -- I just have to figure out how to
> give it, uh, context about the context I want dropped.
I don't actually know...I asked claude if it helped and it said yes, but
claude blows a lot of sunshine about how smart your prompts are.

> 
> And the semcode MCP sounds very interesting! I look forward to finding
> out more about that. :)
> 
> Awesome!
Glad it looks useful, please let me know what bugs you find!

-chris