Boris Cherny on Directing Claude Code Toward Its Own Desktop App
*Anthropic's head of Claude Code says the key skill is assigning hard tasks and building in verification steps rather than refining prompts.*
Boris Cherny, who leads Claude Code at Anthropic, spoke at Y Combinator’s Startup School 2026 about how teams should approach difficult coding jobs with the model. He argued that the main challenge has shifted away from prompt engineering toward selecting tasks that stretch the system and then ensuring it can check its own output.
Cherny noted that verification remains the step most people overlook. He described one internal effort involving the Claude desktop app, which runs on Electron. The team has since improved performance, but six months earlier the experience was noticeably slower.
The example illustrates Cherny’s broader point about giving the model work that initially appears too ambitious and then structuring the process so it can confirm progress at each stage.
Why it matters
Teams that treat verification as an afterthought will continue to see unreliable results even as model capabilities grow. The Electron app case shows that internal groups at Anthropic are applying the same workflow they recommend to others, which may set expectations for how quickly external developers can replicate similar gains.
---
Sources:
No comments yet