Summary
Anthropic’s Claude Code team has changed how it works, going from roughly 200 code changes a day in May to more than 30,000 a day in September. This video from Nate Herk breaks down a talk by Lydia Hi of the Claude Code team and turns it into a setup viewers can copy.
The talk covers five things developers used to do manually in every session. The first is steering: with stronger models like Opus 5.5, emphatic instructions in CLAUDE.md can hurt, so the file should hold only what Claude cannot learn from the code. A /doctor prompt audit helps prune it. The second is approving: auto mode uses a separate classifier to check commands and trusts only the current folder and repo by default, configured in plain English or with /auto mode setup. Anthropic’s study found people caught about 14% of clearly dangerous commands, while auto mode blocked 89%.
The third is saying no, using hard deny rules for anything irreversible, such as reading .env files, sending emails, or deleting files. The video also covers the /sandbox command and disposable cloud sessions for long-running jobs, plus giving Claude pass-or-fail checks such as tests or screenshots so runs that go wrong are caught early. The approach applies to knowledge work as well as software development.
📺 Source: Nate Herk | AI Automation · Published November 17, 2026
🏷️ Format: News Analysis







