Boris Cherny on Trying to Get Claude Code to Rewrite the Claude App

I'll pull the source article and HN thread so the paragraphs stick to verified details only.Boris Cherny, head of Claude Code at Anthropic, described the experiment on stage with Diana Hu at Y Combinator Startup School 2026. He wanted to know how a native Claude desktop would feel after the team had already improved the existing Electron client. Six months earlier that app was sluggish and unreliable; Cherny said the current build is what most of the team uses day to day. As a test, he pointed an agent at a full rewrite instead of a feature patch. The write-up and Hacker News thread (61 points, 13 comments) pulled focus onto that experiment, not a shipping product announcement.

The setup was deliberate tooling, not a single clever prompt. Cherny used Claude Tag, Anthropic’s Claude-in-Slack product, as the control surface. His first question was whether Tag had a macOS runner on GitHub; when it said no, he attached a runner so the agent could boot a Mac virtual machine. He then opened an empty Swift codebase meant to become a native Claude desktop and granted the agent access. The instruction was blunt: rewrite the Electron app in Swift, run the Electron build in the Mac VM, screenshot both, compare pixel by pixel, and do not stop until finished. At the time of the interview the job had been running a little over two weeks—about 14 or 15 days—and was still going.

That framing matters for builders more than the Mac client itself. Cherny’s claim is that the scarce skill is no longer wordsmithing prompts; it is choosing a task that looks slightly too hard and forcing continuous verification while the agent works. Pixel comparison is a crude but concrete loop: visual delta becomes the acceptance test instead of “looks good” judgment. For anyone running long agent sessions, the lesson is architectural: give the model a machine, a repo, a reference binary, and a stop condition that is checkable without a human in the loop every few minutes.

The competitive angle is less about Electron versus Swift and more about who can run multi-week autonomous coding loops with real environment access. Desktop Claude already competes with other chat clients; a native Swift port would only matter if it fixed reliability and feel, not just the framework label. John Gruber’s critique lands here: cloning the current UI pixel-for-pixel can at best transplant layout into AppKit or SwiftUI, not fix non-native design or Mac idioms the Electron app ignores. Pixel parity is a verification strategy for porting chrome, not a product strategy for winning on desktop quality against OpenAI’s client or a well-designed Electron app.

What to watch next is whether the still-running job produces a usable Swift shell, stalls, or gets killed after burning weeks of agent time. Engineers should treat the interview as a verification pattern to copy, not a mandate to burn two weeks on visual diffs. Prefer checks that catch behavior—launch time, crash rates, keyboard paths, accessibility—alongside screenshots, and decide up front how much continuous compute a rewrite is worth before the agent is told not to stop.