Claude Code Thoughts

Summary

I finally had the space to really test Claude Code as a part of my development workflow last week. I was blown away — it’s as good as advertised. For my coding productivity, it’s the biggest step change in productivity in the past year and continues to build my conviction that we’re headed to a world where well defined coding projects are cheap to implement.

The project

My test project was related to this blog. I had chosen SendGrid as an email provider and like a month after I shipped email subscriptions SendGrid got rid of their free tier… and I’m not paying $20 a month to send a handful of emails per week. So I needed to migrate and I had 60 days to do it less I disappoint my adoring public.

I used Claude Code via the Cursor Terminal / Cursor extension. Because of Charlie’s advice, I used Claude Opus for the coding.

To begin with, I gave Claude the background:

• Where my email service currently lives

• Where my email templates currently live

• That I need to move from SendGrid to Resend

Claude Code took it from there — it did a review of my code and concluded this wouldn’t be a particularly impactful change. It reviewed the Resend documentation, outlined the files that needed to change and came back to me with a plan, which I went along with.

Along the way, I skimmed the code changes to make sure they weren’t far reaching in a way that didn’t make sense to me.

The first version of the plan worked but ran into Resend’s rate limits, so we had to switch to the batch version of the API. I knew about the batch version of the API from a previous project, so I could provide it with that link. Without that, it’s possible that the project would’ve gotten stuck there.

All In all, once I was set up, it was incredibly painless — the equivalent of slacking the superstar engineer on the team about a project and suddenly it’s done.

Friction and nits

I ended up using Opus via the API rather than Claude Sonnet via my Claude account. My experience using Opus via Claude Code will probably convince me to set up to the $100 a month plan here soon — I’ll easily hit that in API fees working on Recap this week. But it’s super confusing that my $20 a month plan gives me access to Claude Opus via the Chat App but not via Claude Code and that I have to log out of my account in Claude Code to use my API credits. Just let me stay logged in and use my API credits!

Setting up Claude Code in Cursor was a pain. I had already installed it via the terminal and had to uninstall it and reinstall it via the Cursor command line. For some reason, the Claude extension isn’t available in the extension marketplace any longer, which is confusing to me.

Claude Code’s auto updates have been failing on me ~1x a week and then I have to uninstall and reinstall. No idea why this is but it’s annoying!

Reflections and implications

This was a really comfortable way for me to code. I’m a vibe coder at heart and integrating more deeply into Cursor / Terminal takes a bunch of the copying and pasting out of it. My loop now looks like: 1. Use the Claude App to define what I’m building (using Harper’s scoping prompt) 2. Pull that into a spec (the overall project) and a plan (delivery milestones with a definition of done). Claude writes this and I edit it. I put it in a /prd folder. 3. Have Claude Code read this and create its list. From there we are off.

I think this is going to be a pretty large change for developer tools companies. Previously, if you could get a startup to build around your technology early, you could keep them for a long time simply out of inertia, giving you a fair amount of pricing power. It was expensive to prioritize switching services over feature development and teams would need to feel a fair amount of pain to make a change. Now… not so much? You still have to do the QA of the change, but the coding should be more or less free for well documented services.

Yet again, I’m going to have to re-evaluate what AI procurement. I have been on the ChatGPT pro plan to have access to Deep Research… but it might be time to change that.

2025-07-03