A Claude session has been rewriting an app in Swift for over two weeks
Speaking at Y Combinator’s Startup School 2026, Claude Code lead Boris Cherny said the real skill now isn’t prompt engineering. It’s handing Claude a task slightly beyond what it can comfortably do, and giving the model a way to check its own work as it goes.
His test case: a Claude Tag session (Claude running inside Slack) wired up to spin a macOS virtual machine through GitHub, with access to an empty repo for a Swift build of the desktop app. The prompt was to rewrite the Electron app in Swift, run the Electron version in the Mac VM, take screenshots, compare them pixel by pixel, and not stop until it’s done.
As of the interview, the job had been running a bit over two weeks, 14 to 15 days, and hadn’t finished. In the interview breakdown, launching the current app on a Mac took 30 seconds, the last 15 of them on a spinning wait cursor.
Up next
anthropickit harvested SSH keys during pip install
PyPI package anthropickit ran from setup.py at pip install: it read everything in ~/.ssh, scraped env vars containing KEY/SECRET/TOKEN/PASS/AUTH/API, and posted the haul to a disposable Pipedream endpoint. A PyPI malware blogger calls it a possible match for Anthropic’s Incident 2.
Related stories
- An internal Claude tool now lands 65% of the Claude Code team’s product PRs
- Claude Code team hands most code reading to Claude
- Claude went from bug hunt to abuse report on New Year’s Eve
- Bun’s half-million lines of Zig became Rust in eleven days
- Anthropic’s six-step playbook for migrating code with Claude Code
- Claude recovers 41 TB of BTRFS data
Comments
No comments yet. Be the first.
Join the conversation
Sign in to leave a comment. It appears after moderation.