It’s getting really annoying to review code
Posted by idontwant2makeausern@reddit | ExperiencedDevs | View on Reddit | 23 comments
Whether it’s AI generated entirely or partially, people just don’t know the decisions they’re making anymore and why. I can’t even track what the original ask was to begin with to build the thing let alone why certain changes were made. Obviously this is exacerbated by AI generated slop code. Curious if you’ve found a workflow that works for you.
therealslimshady1234@reddit
Yes, every AI codebase is slowly turning into slop. There is no stopping it it seems
reddit-ate-my-face@reddit
It will stop. Something fucking awful is going to need to happen but not a moment sooner.
therealslimshady1234@reddit
Maybe another AWS outage but this time 1 week instead of a few hours?
ButchDeanCA@reddit
I agree. I’m just patiently waiting for those higher ups that have been pushing AI to realize it was never showtime for this overall nonsense.
Plenty_Line2696@reddit
we have a rule that whoever submits a pr has to screenshare and walk the reviewer through the changes, just high level to ensure they understand/own the changes and prevent slop from reaching production.
itsdikey@reddit
oh I can imagine pushback on this, if anyone suggests something like this.
ExperiencedDevs-ModTeam@reddit
Rule 9: No Low Effort Posts, Excessive Venting, or Bragging.
Using this subreddit to crowd source answers to something that isn't really contributing to the spirit of this subreddit is forbidden at moderator's discretion. This includes posts that are mostly focused around venting or bragging; both of these types of posts are difficult to moderate and don't contribute much to the subreddit.
FireHotTakes@reddit
The only thing worse than ai slop code you have to review, is when your code review is littered with AI generated comments.
I work at Amazon and how much you comment on crs is tracked and sometimes used towards performance. I have a couple devs on my team who have no shame just copy/pasting 8 comments where 6 are just wrong, and the other 2 are just stylistic preference/nits and then I have to spend over an hour responding why I won't make changes for each of them.
EdelinePenrose@reddit
have you suggested standardized ai workflows that include the necessary context for human review?
chizel999@reddit
well, glad you are still reviewing code.
i have a specially annoying coworker who sends me horribly bloated "spec documents" for his "spec driven development workflow" written by AI and actually expects me do read them and offer feedback. lol
idontwant2makeausern@reddit (OP)
I also don’t know why any of these tools don’t solve for this problem. GitHub or Cursor or whatever could conceivably own better translating the diffs, decisions, and even broader technical approach to the reviewer.
throwaway_0x90@reddit
Don't accept large PRs
idontwant2makeausern@reddit (OP)
Totally. Let me tell ya the smaller ones aren’t great either
ribenakifragostafylo@reddit
Also I got a tongue lashing and basically got called a dinosaur when I suggested that size of PRs can be useful to guard against ginormus slop PRs. One person sent me the reddit post about "writing code is dead so will the code review will die too".
throwaway_0x90@reddit
respond to them with this:
throwaway_0x90@reddit
true, but assumably the smaller diffs are much easier to reason about. Combined with the work ticket that hopefully clearly explains what the change is and why.
myaltaccountohyeah@reddit
This was and still is the way. More than 12-15 files changed and I'm not looking at the crap.
Concrete limits obviously vary by codebase, language etc.
MissinqLink@reddit
Everyone quickly agreed to this after the first one.
Devji00@reddit
Completely feel this, the worst part isn't even bad code it's code where you can't reconstruct the reasoning behind any of it. The thing that's helped our team the most is making the PR description non-negotiable, if you used AI to generate the code you still have to write a human explanation of what you asked for, what decisions were made, and what trade-offs you accepted, basically forcing the author to prove they actually understood the output before requesting review. We also hook Cyclopt Companion into our pipeline for security and code quality. But honestly the biggest lever is cultural, we just started rejecting PRs where the author can't explain their own changes in the review thread, and after a few rounds of that people got way more intentional about understanding what they're submitting even when AI wrote most of it.
idontwant2makeausern@reddit (OP)
This is actually a super helpful tip (human explanation of work in PR description)
Devji00@reddit
Yeah it sounds simple but it changed our reviews completely. The trick is being strict about it at first because people will try to get away with "updated auth flow" as the description and that tells you nothing. We ask for three things: what was the goal, what approach did you take and why, and what did you consider but reject. Once that's habit the reviews basically run themselves because half the questions you'd normally ask are already answered before you read a single line of code.
startupwith_jonathan@reddit
Reviewing code nobody remembers writing
ArtSpeaker@reddit
The problem is the lazy devs are shoving the work of "correct" to YOU. Have meetings with them and have THEM explain to YOU what they did. share that pain.