IDEs Are for Micromanagers
Six months ago, if you’d asked me whether a terminal-based coding tool could compete with a full IDE, I would have chortled into my coffee. Six months ago, Cursor (and friends) had it all: slick interfaces, deep integrations, and massive funding. Claude Code was just a TUI. It felt like Anthropic wasn’t being serious about the space and threw a half-baked product into the wild.
Yet, here we are. The above chart is Google Trends for “Cursor” vs “Claude Code”. Claude Code went from essentially zero to overtaking Cursor in about twelve months.
This is the source of much anxiety amongst founders in Silicon Valley right now. You spend years building a product, and then a big lab walks into your market and eats your lunch. The CC vs Cursor story seems to validate that fear perfectly. Anthropic decided to compete, perhaps as an afterthought, and now they’re winning.
But it dawns on me: Claude Code isn’t winning because they have a secret LLM. Cursor uses the same Claude models through the same API. The raw intelligence powering both tools is basically identical. Anthropic does have some structural advantages here: they can fine-tune for their own agentic loops and whatnot, but I don’t think it’s sufficient to explain what’s happening. Cursor is a company laser-focused on developer tools, and you would think they’d be better at building the last mile of polish that would make them win. For example, I’m sure Cursor has some really smart people who think all day about the right system prompt for a coding agent. I’m not sure Anthropic had those resources to spare 6-12 months ago. Cursor is the specialist, Anthropic is the generalist. You’d expect the specialist to win.
And yet I think Claude Code is winning not because of its access to better models, but because of UX. The TUI.
When I first used CC, I thought the minimal interface was a limitation. Now I realize it’s the superpower. Working in CC doesn’t feel like programming. It feels like delegating. I describe what I want at a high level, and the agent executes. I regularly run five CC sessions in parallel across different parts of a codebase. I don’t write code. I provide vision and milestones to the team.
Cursor, for all its polish, still puts me in the programmer’s seat. I’m watching diffs, approving line-by-line changes, tabbing through suggestions. It’s pair programming with a very fast partner. That’s useful, but it’s a fundamentally different mode of work. The IDE, by its nature, keeps you in the weeds. You’re a micromanager with a really good assistant.
The TUI removes all of that. There are no diffs to review inline, no gutters full of suggestions, no tabs competing for your attention. There’s just a conversation and a codebase. That spareness, which I mistook for laziness, actually pushed me into a completely different relationship with my code. I went from line-level thinking to system-level thinking.
This matters for founders because it inverts the conventional fear. The worry is that the big labs will crush startups because they have better models. But CC isn’t winning on models. It’s winning on DevEx. Specifically, it’s winning because someone (perhaps accidentally) discovered that removing features could be more powerful than adding them. The best developer experience innovation of the past year wasn’t some space-age way to navigate your code. It was the revival of a text-based interface that frees you up to think at a higher level.
This suggests the defensible ground for startups isn’t where most founders think it is. If you’re building on top of LLMs and your moat is “we have better prompts” or “we fine-tuned on more data,” you’re competing on the same axis as the labs. They will eventually do it better than you and will crush your soul in the process.
But the labs are big, bureaucratic machines. They are structurally hindered from taking big risks on dimensions like interaction paradigms. They’re not thinking about the weird, counterintuitive product decision that changes how people relate to their tools. Anthropic stumbled (I believe) into this one and I think it’s going to be really hard for them to do it again (innovators dilemma, if you will). The next version of this insight is sitting there waiting for someone to find it.
I think the founders who survive the age of AI labs won’t be the ones with the best model access. It will be the ones with crazy-enough ideas that win on the non-obvious dimensions.