These prompts give AI extra context to more effectively review kernel code. They can be paired with semcode, which makes the review sessions faster and more accurate by indexing the kernel tree, reducing the time AI spends grepping for function/type definitions, and call graphs.
The semcode github repository has instructions on setting up the indexing and MCP server, but if you're just getting started, you can do the quick start without semcode.
Put these prompts somewhere, and then tell claude to use them:
claude> Using the prompt ../review-prompts/review-core.md run a deep dive regression analysis of the top commit
Claude has a internal definition of what "reviewing" code means, so if we call it a review, it will generally follow that internal definition. We can nudge it slightly, but calling it a deep dive regression analysis leads to better compliance with the prompts.
You can also feed it incremental diffs, or use debugging.md with an oops or stack trace.
If you want to use this in GitHub workflows, see this document for integration instructions.
Claude will chat its way through the code review, and that output is pretty useful. If regressions are found, it creates a review-inline.txt file, which is meant to look like an email that would be sent to lkml.
sample.txt has examples of regressions.
Many of the false positives are just AI not understanding the kernel, which is why there are per-subsystem context files. We'll never get down to zero false positives, but the goal is to build up enough knowledge that AI tools can lead us in the right direction.
The false positive rate is improving, currently at ~20-30%.
review-core.md sets the checklist and also tells AI which prompts to conditionally load. Start reading there.
The easiest way to understand a given regression is often to load the regression report into an AI agent and ask questions. The agents are generally accurate at finding details in the kernel tree, so if it claims there is a use-after-free, ask for the call chains or what conditions it might happen. These really help nail things down.
The existing prompts catch a wide variety of bugs, and most subsystems won't need special instructions. If you're finding false positives or missed bugs, it can help to add a few notes to help AI get the review right. patterns/BLOCK-001.md and patterns/LIBBPF-001.md are two examples where we fill in extra details that you can use as a guide.
The basic structure of the prompts continues to change, and should decrease in complexity now that we have a good baseline. But subsystem specific prompts usually just add a few specific details, and can be very short.
technical-patterns.md includes most individual patterns, and review-core.md includes subsystem specific prompts. This lets us limit tokens spent to only the prompts that are relevant to the patch.
Beyond that, the existing prompts are structured to try and make sure AI actually follows the steps. The basic idea:
- Explain when to run this prompt
- Add additional knowledge about code or data structures
- Put a series of steps into a TodoWrite
- Gather context needed to review the code
- Make AI produce output at each step to prove it is following instructions
The BPF CI sends reviews based on these prompts to the BPF mailing list. review-stats.md can be used to compile a report about how effective these reviews are. The idea is to compile a list of all the message-ids from the automated reviews, and then have AI use semcode to pull down those threads and analyze the results one at a time.
It outputs two files, (review-analysis.txt and review-details.txt), and they are meant to be concatenated together after the run is done.
Sample output is in examples/review-stat.txt
Right now I'm more focused on reducing false positives than finding every bug. We need to make sure kernel developers find the output useful and actionable, and then we can start adding it into CI systems.
With that said, patches to firm up any of the subsystem specific details or help it find new classes of bugs are very much appreciated.
The prompts have been developed against claude, but gemini also works well. These should be generic enough that other agents work too, but please send patches if we can improve performance with any of the other agents.
