Ok, end of thread! Hope whoever reads this finds it helpful. I definitely wish CLIs existed when I was a wee third year grad student, so hope others will adopt and benefit :)

Now back to writing the JMP I guess :(

Similarly, the deeper your knowledge of the codebase, the more productively you can interact with Claude in making edits. That’s why I mostly use it to edit or add to my scripts instead of authoring things from scratch.

To emphasize, the better managed your project is, the better it works! If your pipeline is built to be replicable and version controlled on Git, it is way less worrying to have Claude make changes to your code.

Yes, it gets stuff wrong sometimes. But since it’s in your terminal, it can test out scripts locally—a huge plus over browser-based AI!

And it’s easier than ever to iterate with it, e.g. “see how ../Summary/foo.do makes that table? That works properly, and I want the same thing here too”

These are the sorts of use cases—along with debugging—that I’ve personally found Claude Code most useful for. Easily a 5x productivity multiplier in many cases.

The more complex the codebase, the more helpful I’ve found Claude Code to be. For example, I estimate dynamic IO models in Julia. It isn’t always easy to try new specs. Or as the project grows and I want to add functionality, organizing and cleaning up functions is a lot of cognitive overhead.

For example, at the root of my project directory, I wrote a CLAUDE.md doc with a brief style guide and directory org chart. It gets loaded in at the start of every session. Kind of like how you’d want your RA to roughly follow the Gentzkow + Shapiro handbook, or whatever guidelines you use.

As with all AI, it is a complement, not a substitute, for good coding practices like using version control. The complementarities for Claude Code just happen to be very strong.

The integration with git is particularly nice. You can let it see your git log to give it more context, and even have it commit and push changes for you (its commit messages are way better than the jumbled slop I usually write!).

The result? A cool tool that you can use to taste. Give it access to your project directory and it can sift through data pipelines or helper functions across files. No more copy and pasting lines of code to/from your browser—it can edit files directly and highlight the differences for you to okay.

I can choose which directories it can see, it won’t run terminal commands unless I say so, and it never makes edits without my approval. (You can “yes to all” each session if you want it to.)

And ofc it's a CLI, so you can call it directly from your terminal. This is huge for vim guys like me!

The main point: it lives on your machine so you can call it straight from the terminal. But it always, *always* asks for permission!

I’ve found that this massively lowers the barrier to entry that previously existed for opening an AI chat before, but keeps me assured that I’m using something safe.

Disclaimer: I hadn’t previously used anything fancier than ChatGPT/Claude in the web browser. A lot of benefits I’m about to mention already existed in tools like Cursor or Github Copilot--though there are still some features that I do think are unique to Claude Code!