Stack Check, March '26

March 5, 2026 · 1119 words · 6 min read

How my dev tools finally stopped fighting each other.

Oof, here I go again, I've written about this twice now. First in The Joy of Keeping Your Hands on the Tools, where I shared how I finally figured out that AI-assisted coding didn't threaten the craft I cared about, and then in Staying Close, Moving Faster, where I discussed finding a rhythm that lets me draft fast and decide slow. This is the update nobody asked for, but I'm writing it anyway because my stack changed and the reasons feel blabbing about.

The short version is that I moved to Claude Code on the Max plan as my primary tool, went back to VS Code with GitHub Copilot for editor work, and stopped using Cursor entirely. The longer version involves some frustration, some planning, and my continued conviction that the best workflow isn't the one with the most impressive tools in it, it's actually the one where everything stays out of your way and shuts the hell up when you don't need it.

I'll start with what I left behind. Cursor's tab completion was, and probably still is, the best in class. After they acquired Supermaven in late 2024, the completions got even sharper and faster, I have built strong muscle memory around it. Tab, tab, tab through long edits, and the suggestions felt like they were reading my mind! That part was absolutely great and I miss it dearly.

But everything else decayed, the UI started feeling cluttered and unreliable, extensions would break in ways that VS Code extensions simply don't, because Cursor is a fork that drifts further from upstream with every update, and the cracks show up in the strangest places. The extension registry felt increasingly like a bazaar of CVEs and half the things you installed introduced subtle weirdness you'd only notice three days later when something stopped working. I don't want to spend my limited evening coding time debugging my editor, I want the editor to disappear and let me work and do what I love.

And then they sunsetted Supermaven entirely in November 2025. I understand the business logic and I hold nothing against them, they absorbed the team and folded the technology into Cursor's own models. But it still stung! Supermaven was one of the few pure code completion products that did one thing incredibly well, its flexibility and competence mattered to me, and watching it get absorbed into a single editor's ecosystem felt like losing a tool that respected how I actually work. What frustrates me more broadly is that pure code completion as a standalone product category seems to be dying. Everything wants to be an agent now, everything wants to own the whole experience, and the simple "just finish my line intelligently" use case keeps getting bundled into bigger, more opinionated packages. I liked having a sharp knife, especially when everyone wants to sell me the whole kitchen.

So I went back to VS Code with the Copilot plan. It's a tradeoff, the completion quality isn't as good as what Cursor had with Supermaven under the hood aand while Copilot's suggestions are fine, even they're competent, they still land maybe 60% of the time where Supermaven landed 85%. But at least the UX is stable, the extension ecosystem is battle-tested, and I'm not fighting the editor itself. That stability tax is one I'm happy to pay because when I'm in the editor, I want to be thinking about code, not about why my sidebar just rearranged itself.

Claude Code is where the real shift happened though. I mentioned in Part 2 that I used it for async work, but back then it was one tool among several, now it's the center of gravity! The Max plan makes the economics work for how I use it and I only hit Pro limits within the first week of trying it, once I moved to Max the rate limits stopped being something I thought about during a normal evening session. It's genuinely worth the money if you use it the way I do.

What makes Claude Code different from everything else I've tried is that it operates at the right level of abstraction for how I think about problems. It's also not trying to be an editor or autocompleting my keystrokes. It's an agent that lives in the terminal, understands my project structure, reads my files, runs my tests, and does real multi-step work while I go make coffee or review something else. The async pattern from Part 2, "go do this and show me what you did," is now my default mode for a huge category of work. I still have to set intent and constraints ruthlessly, and it still mostly does the mechanical work while I review the diff, but it replaced several tools at once and simplified my setup dramatically.

But the thing I didn't expect was how well it works for the close, hands-on work too. When I'm stuck on a design question or trying to figure out the right data flow for something new, I can think out loud with it in the terminal. I can paste an image and show the problem, and while it's not the same as having a colleague look over your shoulder, it's closer to that feeling than anything else I've used. I describe the problem, it proposes a direction, I push back, it revises, and within a few rounds I usually have something I can work with in the editor. The conversation stays in the terminal, the code stays in my files, and the boundary between "thinking" and "building" gets thinner without disappearing entirely.

The whole setup now feels like it reflects the cadence I need. VS Code with Copilot handles the slow, precise, hands-on-the-keyboard work. They don't overlap, they don't compete, and more importantly neither one tries to be the other. That separation is the thing that took me three posts and over a year to arrive at, and it's embarrassingly simple in hindsight. The best tool for exploring cannot be the best tool for refining and the best tool for delegating cannot be the best tool for deciding. Trying to make one tool do everything is how you end up with a bloated editor that fights you, or an agent that doesn't know when to stop.

I'm also more at peace with the fact that this will change again. Whatever I'm using six months from now will probably be different, and that's fine. The workflow is the thing that stabilized, not the tools. Draft fast, stay close for the parts that matter, own what you ship. The tools are just how that workflow expresses itself on any given Tuesday, and right now, this particular Tuesday feels pretty good.