Corsair

In the summer of 2024, the bottleneck was schlepping text between browser and editor.

I was using Emacs all day, moving things around in text buffers, talking to Claude (Opus 3 at the time) via gptel and ChatGPT in the browser. The workflow was incremental: paste hundreds of lines of code, talk through approaches, ask about tests, sketch out a new class. Very controlled. Very manual.

I built Corsair to smooth this out - a small Emacs package for sending code to LLMs and yanking responses back into buffers. Reducing the friction for schlepping.

I was happy to schlep. It was faster if you did it carefully, and I could tell the tools were growing. Worth learning.

Already Obsolete

The harnesses have since overwhelmed my ability to build tools.

Corsair is no longer something I use. You can now talk directly to an agent that sees your code, edits it, runs tests, iterates. Even copy-pasting has been obviated - in the rare case I need something in my system clipboard, I just ask.

The bottleneck moved.

Where It Is Now

At the moment, I’d say the bottleneck is voice input. Getting intent across. Communicating the latent context around a project - what I’m trying to achieve, what constraints matter, what I’ve already tried.

Dwarkesh Patel has a post arguing that continual learning is the next big bottleneck. Models don’t get better over time the way humans do. They can’t yet accumulate project-specific knowledge across sessions.

When that changes, my workflow will change again. Until then, it might stay roughly stable.

The Lesson

I don’t regret building Corsair. Scratching your own itch is how you learn. The schlepping taught me how to work with these systems. The tool was a symptom of engagement, not wasted effort.

But it’s a clean example of how fast things move. Less than a year from useful to obsolete. Not because it broke - because the world moved past it.