IT/Software career thread: Invert binary trees for dollars.

Noodleface

Noodleface

A Mod Real Quick
39,878
18,651
Deathwing said:
Does anyone here use AI for reviewing code?
clear conversation -> Review -> Fix or refute(and save to memory!) -> repeat
I've never been able to get it to converge, it always finds new shit. And later into that cycle, a significant chunk of the findings are code changes it made.
Click to expand...
Not my own code, but I use it for other's PRs.

I basically point my agent at their branch and ask it to do a code review on it (and provide no context). I then look at what it recommends and verify for myself. Anything I think is worthy, I comment in the PR or ask questions about.

What I HATE is this guy on my team that does it and just copies and pastes the diarrhea Claude spits out (like 10 paragraphs worth) with no human interpretation behind it. Like I said, this is how old heads are using it. It's essentially useless and I always ignore it if people do that.
 
  • 2Like
  • 1Truth!
Reactions: 2 users
Deathwing

Deathwing

<Bronze Donator>
17,801
8,760
Control said:
Only hobby code, so maybe not too relevant, but I've been having one model orchestrate another to check and another to fix so they basically argue amongst themselves until the orchestrator is satisfied. They eventually sort it out, but it takes forever. Generally Fable to orchestrate and swapping between Sol/GLM to review/fix. I mostly did it that way to budget/spread tokens, and it works, but probably mostly because it takes so damned long.
Click to expand...
That's definitely a "solution" and one my boss went to almost immediately. He's on the $200 Claude Code plan and was complaining about running out of tokens 5 days into his weekly session. I was lamenting having to ask for more(because I don't think we've solved the baggage that comes with AI) just to participate in whatever clown car of a workflow my boss is cooking up. Now I'm thinking I have more immediate counter points. Spending 90% of your feature in MR review is demoralizing.
 
Deathwing

Deathwing

<Bronze Donator>
17,801
8,760
Noodleface said:
Not my own code, but I use it for other's PRs.

I basically point my agent at their branch and ask it to do a code review on it (and provide no context). I then look at what it recommends and verify for myself. Anything I think is worthy, I comment in the PR or ask questions about.

What I HATE is this guy on my team that does it and just copies and pastes the diarrhea Claude spits out (like 10 paragraphs worth) with no human interpretation behind it. Like I said, this is how old heads are using it. It's essentially useless and I always ignore it if people do that.
Click to expand...
Can you tell the difference if someone had already done an AI review on their code? If not, doesn't that imply you should review your own code similarly? If you can, do you toss the review away as it's potentially already into garbage territory?

I've definitely done what your teammate has done. Partly because it's easy and partly because my boss was doing it, so...I guess that's what we do now. Thanks for reminding me, I'm going to stop doing that, I like your methodology better.

Claude's verboseness has definitely been an issue on my team. Some people don't like how much reading it entails. If you don't put any leashes on it, the MR descriptions, commit messages, thread comments, and code comments can be gigantic.

Another problem is that it can seem insincere. During a team discussion on this topic, one of my MRs was used as an example. My boss started 9 threads, each at least 4 paragraphs long, took me all afternoon to respond and fix those threads. People thought the review was insincere because it was obviously written by AI. My boss said he spent 4 hours chatting with the AI before filing each thread.
WTF2.png
 
  • 1WTF
  • 1Like
Reactions: 1 users
Cad

Sheriff Cad

scientia potentia est
<Nazi Janitors>
34,880
85,281
Deathwing said:
That's definitely a "solution" and one my boss went to almost immediately. He's on the $200 Claude Code plan and was complaining about running out of tokens 5 days into his weekly session. I was lamenting having to ask for more(because I don't think we've solved the baggage that comes with AI) just to participate in whatever clown car of a workflow my boss is cooking up. Now I'm thinking I have more immediate counter points. Spending 90% of your feature in MR review is demoralizing.
Click to expand...
Haven't coded at all with AI but isn't this what the "agents" do in multi-agent plans? Grok used to have the multi-agent thing a few months ago with their $30 plan, but now they hid it behind the $100 plan. It would show the agents asking each other questions, giving each other tasks, checking on the results of this or that, and I thought that would be a really smart way to avoid some of the dumb AI pitfalls, have a supervisor model that asks other models questions and sanity-checks the answer. Whenever I draft a brief using AI, I always polish it up to where I'm ready to file, then run the whole thing again through a fresh conversation for final review, and it "often" finds little things that the other conversation didn't - same model, exact same brief.

Anyway, is the multi-agent thing doing anything like what we're talking about?
 
Control

Control

Golden Baronet of the Realm
6,057
16,895
Cad said:
Anyway, is the multi-agent thing doing anything like what we're talking about?
Click to expand...
It can if instructed, but that's not necessarily the default behavior. Fable may have changed this a bit (or it may have been harness updates... I think ultracode does it by default), but previously, if I just told claude to spawn some agents to do x without any other instructions on tests or verification, it would mostly take their word for it that they were done. I would specifically define review agents and tell it to fire them off afterwards, results could be pretty hit or miss. My recent workflow has been to tell fable to only act as an orchestrator and to dispatch specific models to implement and review. If anything in the review is flagged, it gets kicked back to the implementer, and the orchestrator doesn't allow commits until the review comes back clean. They seem to stick to the plan pretty well, but how much all of that changes the overall quality of the output or the efficiency is an entirely different question.

Also, for whatever it's worth, this is coming from the perspective of not giving a shit what the code looks like as long as it works, so hopefully that's not how it's being used in organizations (but it probably still is more often than not).
 
Noodleface

Noodleface

A Mod Real Quick
39,878
18,651
Deathwing said:
Can you tell the difference if someone had already done an AI review on their code? If not, doesn't that imply you should review your own code similarly? If you can, do you toss the review away as it's potentially already into garbage territory?

I've definitely done what your teammate has done. Partly because it's easy and partly because my boss was doing it, so...I guess that's what we do now. Thanks for reminding me, I'm going to stop doing that, I like your methodology better.

Claude's verboseness has definitely been an issue on my team. Some people don't like how much reading it entails. If you don't put any leashes on it, the MR descriptions, commit messages, thread comments, and code comments can be gigantic.

Another problem is that it can seem insincere. During a team discussion on this topic, one of my MRs was used as an example. My boss started 9 threads, each at least 4 paragraphs long, took me all afternoon to respond and fix those threads. People thought the review was insincere because it was obviously written by AI. My boss said he spent 4 hours chatting with the AI before filing each thread.
WTF2.png
Click to expand...
I suppose you're right that I could have another agent review my code too. Something to at least try out and see what results I get.

I do think AI is valuable in code reviews, I just think people are way too reliant on AI and spit out what it says verbatim.
 
ShakyJake

ShakyJake

<Donor>
8,772
21,546
Noodleface said:
I suppose you're right that I could have another agent review my code too. Something to at least try out and see what results I get.

I do think AI is valuable in code reviews, I just think people are way too reliant on AI and spit out what it says verbatim.
Click to expand...
I have the agent look over a PR and write its findings to a separate markdown file that I review myself. It can get pretty nitpicky, so I just pull out the things that are legit concerns.
 
Noodleface

Noodleface

A Mod Real Quick
39,878
18,651
ShakyJake said:
I have the agent look over a PR and write its findings to a separate markdown file that I review myself. It can get pretty nitpicky, so I just pull out the things that are legit concerns.
Click to expand...
This is basically how I do PRs and what id advise others do too. People just dumping the output into a PR drives me insane.
 
You must log in or register to reply here.