AI code review (Claude, maybe Codex)

[Eric]

On Saturday, April 11th, 2026 at 16:22, Hauke Mehrtens <hauke at hauke-m.de> wrote:
> On 4/9/26 20:04, Eric wrote:
> > After reading various AI generated reviews, I'd say my only functional objection is that they are so f'ing verbose.
> >
> > If I (speaking as a human reviewer) see code like this in a PR:
> >
> >     cp $file $dest
> >
> > I just say "Please quote the variables to avoid issues with blanks, wildcard expansions, etc (do you know about 'shellcheck'?)", assuming a high level of comprehension on the submitter.  When an AI sees that, it appears to assume the coder has the expertise of a 3-year-old, so produces a PhD dissertation as to why you should add quotes.  Reviewing the review becomes 10x more burdensome than it should.
> 
> Hi,
> 
> I played around with the Claude code github action again and added an
> additional system prompt to do less verbose messages. They should also
> be cheaper. ;-)
> 
> It now also should do inline comments. See here my test review:
> https://github.com/hauke/openwrt/pull/10
> 
> The inline comments are also nice, you can answer on each of them and
> then mark then as resolved.

Agreed, those inline comments are much better than the "one
giant comment", which makes you jump around so much.  I may
be hallucinating this, but it seems like the inline comments
are more on-point, too, less rambling about justifications and
less "fluff".

Eric