Working with a Massive Codebase/ What Millions of Lines of Code MeanLESSON 4 OF 21Log inSubscribe
Repomix and Graphify lesson slide
LESSON 4 OF 21 · IN THE FULL COURSE

This one's part of the full course

Good to see you. This one's in the full course

2 lessons are free to watch right now, no account needed. If they click, grab the course and every lesson is yours to keep.

You've seen how I teach. The other 19 lessons are in the full course, and your progress comes with you.

WATCH THESE FREE, NO ACCOUNT NEEDEDYOUR FREE LESSONS

Already bought it? Log in and it picks up right here.

Your progress comes with you the moment you buy.

What Millions of Lines of Code Mean · LESSON 4

Repomix and Graphify

THE COURSE, END TO ENDNow playingDoneAheadYou are hereFree to watchIn the full course
What Millions of Lines of Code Mean
Making the Codebase Yours
Ripping Out Copilot
Rethinking the Editor
ABOUT THIS LESSON

Time to map the repo, and the rule still holds: I am not opening directories.

There are two popular tools for this and I try both.

Repomix, from the CodeRabbit folks, crawls your project and emits one “AI friendly” file. I have used it before and liked it. So I run it here. It is fast. 16,446 files. Then I open the output.

Five million lines. 200MB of XML.

I take it to Claude anyway, with a fresh context, and ask straight out: can you use this?

  • 50 to 60 million tokens.
  • My context window is one million.
  • No. Not even a single meaningful chunk without blowing the window.
  • Grepping it works, but grepping it is strictly worse than grepping the actual repo.

Repomix is the wrong tool for this repo. Not a bad tool. The wrong one. Its whole model is dump everything and let the model sort it out, and at this scale there is no model that can sort it out.

Graphify is the other option, and it is the number three repo of the day on GitHub, so people are talking about it. It extracts structure rather than dumping content: breadth-first and depth-first traversal, an HTML report, a graph.json of relationships between files. You read the traversal result, not the codebase. It even marks ambiguous edges so you know what the AST could not resolve. That is a genuinely better idea.

Except its own guard trips at 500 files. VS Code is just past what it can take, and it will stop and make you pick a subfolder.

So both are out, and I delete the Repomix output.

But the failure taught me the real question. Do I need to know what is where everywhere, or only for the part I want to change? I like the editing experience. I am not rewriting it. What I want is the file explorer, the shell, some diff functionality, and room for the visual ideas I have.

I do not need a map of VS Code. I need a map of my problem.

So next: a CLAUDE.md with rules about being terse, and a targeted map. Which means I have to tell Claude what I am actually up to.

21lessons across 4 sections
2free to watch right now
Foreverbuy it once, it's yours for life

21 lessons · 2.0 hours · all done

You finished Working with a Massive Codebase.

Every lesson, start to finish. That takes some doing, and you did it. Now go build something with it.

Cheers! Rob