Setup

Claude Code

While Anthropic released Claude Code and Sonnet 3.7 in the spring, the Claude 4 release is what actually allowed me to completely change the way I work. Instead of opening multiple copies of Cursor, I now have multiple copies of Claude Code instead.

My go-to workflow is to use Opus 4 to plan a project, then let Sonnet 4 generate it, and then use Opus again to review the work done. The additions of a discrete planning mode, improved slash commands, and the opusplan model over the course of the summer made this even easier.

While I also prefer the text-based UI of claude to Cursor, I switched away because Cursor’s context window felt very small. When working on a list of 20 items, Cursor would autocomplete 8-10 at a time and then I’d end up with two halves of the list looking subtly different. Claude Code, on the other hand, one shot the whole thing.

Having “full” access to the context felt like a breath of fresh air. It let me work on larger problems, and allowed me to work on multiple problems in parallel since I could trust the model to do more on its own unsupervised.

With that freedom, I’ve ended up with a three “track” workspace, where I have three different tabs open with three copies of claude each working on different problems. One track is for small, quick problems. The second is for larger problems, usually what I’m actually working on at any given time. The last track is for larger, exploratory work that may or may not pan out.

The downside is that, unlike Cursor’s $20/mo, the company’s Teams plan means we are paying for every token. It’s pretty easy to burn through $100/day of tokens (I’ve even used $100 in less than an hour before) and I found myself spending $1200-1600/mo on the company plan, split relatively evenly between Sonnet and Opus.

That’s why I decided to also get the personal Pro Max 200 plan. Even after the late August nerf, I feel like I get considerably more value out of the subscription. Since getting the plan I’ve effectively halved my spend, and shifted the balance on the plan closer to 2/3 Sonnet and 1/3 Opus. Meanwhile, I shifted the exploratory work to the subscription, capped at a manageable $200, while still 3-5x that in value.

Deep Research

Some of my coworkers still really like Deep Research. I tried to use it to plan a few trips over the course of the summer, and it turns out travel agents still have a job.

My real issue with it is how those coworkers use it. They tend to just drop links in Slack to a multipage document they themselves either didn’t read, didn’t verify, or don’t understand. We’ve made multiple missteps this summer due to this. And bluntly, nobody wants to read a 20 page report that may or may not be correct. Wouldn’t it be better to just synthesize the results in your own words and then link to the actual sources for the deep research as justification?

That said, I still do like the report format for my own consumption.

Devin

Some of my coworkers also like Devin, but I almost never use it. We’ve set up Devin with our dev environment, but it’s a bit finicky (just like our dev environment). The real benefit of Devin is not having to do the work locally. One of my coworkers had 8 different small tasks going in parallel in Devin at once. While Claude Code can do something similar with git sub-trees, setting up Docker per environment is tedious. If I use Devin, it’s usually to pick up a small task directly from Slack or our ticketing system.

Cursor

I now really only open Cursor because I need a text editor, not because it is Cursor. I know that Cursor released a bunch of cool features in the 1.0 release and since, but I find most of those features are just tedious and get in my way.