Michael Nuñez reports on VentureBeat on the viral impact of an X thread by Boris Cherny, creator and head of Claude Code at Anthropic, detailing his development workflow. The developer community describes these revelations as a pivotal moment for Anthropic.
Cherny's workflow relies on running five Claude agents in parallel in his terminal. He numbers his iTerm2 tabs from 1 to 5 and uses system notifications to know when an agent requires intervention. While one agent runs a test suite, another refactors a legacy module, and a third writes documentation. He also launches 5 to 10 additional Claude sessions on claude.ai, using a "teleport" command to transfer sessions between the web and his local machine.
Unlike the industry's obsession with latency, Cherny exclusively uses Opus 4.5, Anthropic's heaviest and slowest model. His justification: although slower, this model requires fewer corrections and excels at tool use, ultimately making it faster than a lightweight model. The bottleneck is not generation speed but the human time spent correcting the AI's errors.
To solve the problem of LLM amnesia between sessions, the team maintains a single CLAUDE.md file in their git repository. Every time Claude makes a mistake, it is documented in this file to prevent its repetition. This practice transforms the codebase into a self-correcting organism where every error becomes a permanent rule.
Cherny uses slash commands - custom shortcuts versioned in the repository - to automate repetitive tasks. His /commit-push-pr command, used dozens of times a day, automatically handles versioning. He also deploys specialized subagents: a code-simplifier to clean up architecture and a verify-app for end-to-end testing.
The verification loop is the key innovation. Claude tests each change via the Chrome extension, opens a browser, tests the interface, and iterates until the code works and the UX is satisfactory. This automatic verification improves final quality by 2-3x.
The article concludes with a paradigm shift: AI is no longer an autocomplete assistant but an "operating system for work itself." Developers who adopt this vision will not just be more productive - they will be playing an entirely different game.