Imagine opening your email, seeing a new PR, clicking the link and seeing this:
A wonderful 2300+ line pull request
I suspect many of you would have a small laugh and close the review straight away. Some might fire back a reply and ask for smaller PRs for review, or others might shrug and let Claude do the initial review for you.
I'm not judging any of these responses because I don't think that we, as a community, have caught up with what AI tools can do for us.
This PR is a single sprint that was completed in about 30 minutes. The first iteration of a Stripe webhook receiver that I would easily consider an MVP; it has generous test coverage, and code that I think is pretty dang good, given the task.
Yet the PR is 2300+ lines! How is anyone not using AI supposed to review this? And the one that's behind it, that will likely be even bigger and landing in the next hour!
It's Me, Hi, I'm the Bottleneck It's Me
A human should be reviewing any and all code that goes into production; I think we can agree on this. But when, how, and where is this supposed to happen?
What if this guy was on your team?
People put the wildest things on LinkedIn. To be clear I understand the point being made: don't make me skim your review. But if we know that AI generated the code in the PR, do we need to review it in the same way as before?
I don't think so, but I also know we're adjusting to this reality and we're going to have to figure this out. Throwing more AI at an AI-generated problem might not be the solution we want, but slowing the progress down is definitely not the solution either.
Especially when it comes to the Big Enterprise Refactors coming our way. I do have a solution for this, I think, so read on...
2000 is Cool, But You Know What's Even Cooler...?
I'm sure many of you have worked on the giant Enterprise App at some point in your past, or are working on it now. Imagine being asked by your lead to "have Claude take a look at it" to see how you could upgrade, refactor, fix a bug, add a feature, or do basically anything?
In an ideal world you would work on things in small chunks. In the Not-So-Ideal world we all live in, it's likely that a simple rename will touch 100+ files and result in a gigantic PR, all the time.
How are we supposed to handle these situations?
Well, friendos, that's why you're here cuz ya'boy likes to explore and break things, and that's exactly what I have in queue, and what I'll be shipping this week.
My Billion Dollar Move
I have an idea and I'm going to do it this time. Rather than be all secretive and silly, I'm going to build it in public and see what happens. It might be great and I'll make tons of money and spend my time giving it away to help others, or I'll fail miserably and still help others as they watch how and why I crashed and burned.
I'm forking VS Code and I'm going to build the editor that I wished I had. I'm going to rip out Copilot completely (and a few other things) and then I'm going to implement some features that I wish VS Code had.
I'll be recording this, as I do, editing for clarity and time, and this time I'm going to make it a bit more polished and focused. I'm not just going to mumble my way through as I encounter problems - I'm going to edit, enhance, add some post-production bits and b-roll too. I want to make this as condensed and watchable as I can, because the point of the video isn't going to be me and my billion dollar idea, it's how to work with a massive code base.
VS Code, if you didn't know, has millions of lines of code and is quite complex. Copilot is sprinkled throughout, so this task is not a simple one. I'm going to have to juggle:
Weekly and session limits, which means I need to think about agent use and which model does what.
Token spend, overall. I'm using my Claude Max sub, but I know many of you are on team plan, so we'll keep tabs on the spend, and how to avoid a mess.
Context rot, because working with loads of files across a massive project is prime for blowing up your session window.
This isn't a joke or a marketing play: there's real money on the line here if I do this right (I hope), and I might honestly fail.
Coming in Drops
I'm going to be recording and dropping each episode as I can, but I'm hoping to have this all finished up in the next few monts. It's ambitious, but I have a schedule laid out and I think I can hit it - something I wouldn't have ever thought possible if I didn't have Claude.
Working with large systems isn't a thing that people have been doing for a long time, so I'll be making it up a bit as I go. I've already done some deep research on this, but the new tool additions and larger context windows of the newer models makes it a bit more straightforward.
If you want to follow along with me, I'll be posting the things I learn in my Field Notes space on AI Pro, including some ideas about how you can handle 2000+ line PR. I'll add some videos there too, and the course will end up somewhere on the site, available to premium folks.
Drop by and say hi - joining is free and I have a load of resources you can access just by joining up. No spam or garbage posts either, just you and a thousand or so of senior programmers just like you, hoping to learn more about AI.
If you had any suggestions, hit reply and let me know! Also: I'll publish the course outline in the coming week, and you can add your thoughts there, if you like, as well.
Thanks for reading! Yes, I wrote this by hand...
Rob
P.S. - I have a PR skill that goes through large PRs like the one above and triages them for me using the GitHub CLI. It also gives me a summary of the changes, links to issues I should add comments too, and more. I'll be publishing that skill and how I use it in the coming week.