Inheriting a massive codebase, knowing what's there

I just pushed 7 more lessons to The Massive Codebase, which is the latest course I’m working on. I have a plan, but I’m literally “doing it live”, which means that my plans and ideas could absolutely fail hard.

Speaking of fail hard: I’m sending this to you from my new list mailer app, Koa, which is absolutely blowing my mind. It’s gorgeous and I wove in all kinds of sales and order information, including templating for thank you emails and more. If you notice anything weird, please tell me!

Anyway: I’m working with the VS Code codebase, which is many megabytes of very old code. Millions and millions of lines of code, written by some very talented people.

I want to rip out Copilot and all the AI “stuff” in there, and replace it with my own ideas. I think I have some good ones! We’ll see if that’s true, but first I need to figure out how to:

• Map out what’s there.

• Understand how it works.

• Investigate the history of the project to see where they’ve been, and where they want to go.

• Understand the meta artifacts in there (the .github and AI stuff) and if I can use them.

Yeah, that’s a lot of work… how do you even get started?

Knowing What You Don’t Know

The first step is to create a map of the codebase, and currently there are three “main” ways I’ve discovered to do just this:

• Use a tool like repomix or graphify.

• Use Anthropic’s /codemap skill.

• Do my own thing.

Normally I would never suggest you do the last item there because someone else has already thought through this problem! It’s not a very complex task, so let them do it!

But then again, we’re working with a gigantic codebase, which changes everything because:

• Both graphify and repomix produce artifacts that are hundreds of megabytes - an absolute no go!

• The /codemap skill works, but produces a ton of extra code that we really don’t need.

Long story short: both repomix and graphify didn’t work for me, and you can see me use them (and the results of that) right here, for free.

When it comes down to it: I have a simple need: just tell me where the Copilot stuff is so I can rip it out. That means my best option is to create my own map.

Rolling My Own Mapper

We know the problem: we have a ton of files and we need to know what they do so we can potentially remove them. That’s going to be a very tough problem to tackle, so our map needs to be highly specialized.

But does it require a lot of “thinking” from our model? I don’t think so.

The Haiku Experiment

I tried this a few months ago with Haiku, Claude’s least capable model. It’s good for simple tasks like Git commit messages and creating a data access layer from a SQL file using a template - there’s just not a lot of thinking there.

But what about file mapping?

I ran Haiku over VS Code’s codebase, and then I asked Opus 4.7 how it did, and long story short it didn’t do so well, hallucinating both files and paths. That said, Opus did suggest that it was, overall, a good use of tokens because Haiku did a lot of things right, and checking that is easy for Opus.

The SONNET EXPERIMENT

If Haiku did a good job, what about Sonnet 5, which is far more capable and also pretty cheap these days?

I ran the same experiment, but this time had Opus 5 check Sonnet’s work. Surprisingly, Opus told me not to trust Sonnet’s work!

This was… unexpected! Here’s the full rundown of what Sonnet did when mapping VS Code’s codebase, specific to removing Copilot (it earned a B grade):

Just Spend the Tokens

The map is going to be the most important thing I have in this entire effort because it will save me the most tokens, and keep work focused. Why worry about the token spend on something that important!

So I just told Opus to crack together the map for me, which made the most sense.

Want to follow along? The course is live and I’m adding a few episodes per week as I go along. I’m thinking, overall, about 30 episodes (or “lessons”, I guess, but I don’t really like that word).

How to Work with a Massive Codebase

Thanks for reading this inaugural post from my new, favorite app!

Rob

P.S. This is Koa’s interface. If you’re interested in how I made it, join the Discord and let me know! I might sell it, since it’s highly specialized. I don’t know yet. If you’re a content person and have been looking for an alternative to Kit/Mailchimp - this is for you.

It runs on Cloudflare, has an MCP server built in, and sends with Resend (costs $20/mo) but you can swap that out for Amazon SES.

I love building things with Claude!