Anthropic Engineer Describes Task Assignment Over Prompt Engineering for Claude
*Head of Claude Code Boris Cherny said verification steps matter more than clever wording when directing the model on difficult work.*
Boris Cherny, who leads Claude Code at Anthropic, spoke last week at Y Combinator’s Startup School 2026. In an interview with Diana Hu, he argued that effective use of the model now centers on handing it tasks that initially appear too demanding and then requiring it to check its own output at each stage.
Cherny stated that prompt engineering has become less central. The larger gap, in his view, is teaching the model to verify progress along the way. He described verification as the element most users still overlook.
One concrete case he cited involved Anthropic’s own desktop application. The app is built with Electron. Six months earlier it felt sluggish; recent changes have made it noticeably faster and more responsive.
The interview setting
The remarks began around the twenty-minute mark of the recorded session. Cherny framed the discussion around internal experiments in which the team asked Claude to perform substantial refactors on production code, including parts of the Claude desktop client itself.
He did not provide further technical metrics or timelines beyond the Electron example.
Reactions
No other speakers or follow-up statements appear in the available account of the talk.
Why it matters
Teams that treat model output as a first draft and then layer explicit checks see more reliable results on large changes. The Electron app anecdote shows the difference between an early, slow build and a later polished release; the same pattern can apply when the model is asked to edit its own surrounding codebase. Developers who skip verification steps will continue to hit the same friction Cherny observed.
The approach also shifts hiring and workflow priorities inside AI labs. Instead of searching for the perfect prompt, engineers now spend more time defining intermediate success criteria that the model can test against.
---
Sources:
No comments yet