> When I use AI to review a PR, I prefer to read its findings (you feel the
pain immediately), rephrase the valid points in my own words, and post them
manually.

+100

You absolutely must read and understand your LLM-generated output. I do not
recommend ever posting it directly. It's fine to substantially reuse
wording, but I think it's pretty rare to have a generated comment that
should be directly copied.

I've gone through a lot of generated code and comments in the last year and
it's pretty clear when a contributor has not taken the time to improve on
the LLM output. For PR comments or emails to the list, it tends to be much
longer than needed and overly specific (in many cases, I can debug what you
asked for). For code, I see a lot of really bad tests, useless comments,
Javadoc that leaks implementation details, and naively direct changes.

When I see the signs that something was generated but not edited or
improved, that's a big red flag. I can type "generate tests for
SomeJavaClass" into a chatbot. That's not really contributing. A meaningful
contribution is thinking about what should be tested and how, then
generating those tests and making sure they do what you think they should.
You can do that by having a chat with an LLM about how to produce high
quality tests for a component. Just verify that it makes sense and iterate
to make it better if it doesn't.

I think Russell's Law is a good rule: you need to do the first review of
what an LLM produces, not the project's committers.



On Sat, Aug 15, 2026 at 8:12 AM Manu Zhang <[email protected]> wrote:

> Hi all,
>
> I believe most of us are now using AI to author and review PRs. While it
> is fast and efficient, I have a few concerns regarding how we use it for
> code reviews.
>
> Recently, I have noticed increasingly lengthy AI-generated comments.
> Dissecting these comments to find the root issue takes significant time.
> Why not hand over the job to AI? Yes, I do this all the time but I still
> have to review its proposed action plan. If that plan doesn't make sense, I
> must trace back to the original review. Furthermore, conflicting comments
> from different AI reviewers can cause an inefficient loop, as each tool
> operates on different context. I'm not sure human reviewers ever review
> these AI comments before posting them. I'm also unsure this process remains
> cost-effective given the token expenses.
>
> When I use AI to review a PR, I prefer to read its findings (you feel the
> pain immediately), rephrase the valid points in my own words, and post them
> manually.
>
> Would this approach be a better way for us to utilize AI reviewers? I
> would appreciate your thoughts.
>
>
> *Disclosure: This email was polished by AI.*
> Best regards,
> Manu
>
>

Reply via email to