The Workflow That Broke Developer Twitter

Boris Cherny, the Anthropic engineer who built Claude Code, pulled back the curtain on his day-to-day development workflow in early January 2026, and the programming community has not stopped talking about it since. His approach, which leans heavily on parallel AI agents, custom slash commands, and a living knowledge file, represents a paradigm shift in how software gets written at one of the world's most influential AI companies.

Cherny's workflow is built around a concept he calls subagents: specialized AI personas that handle discrete phases of the development lifecycle. Rather than treating Claude Code as a single monolithic assistant, Cherny deploys dedicated agents for different tasks. A code-simplifier agent handles architectural cleanup. A verify-app agent runs end-to-end tests. Each operates independently, allowing Cherny to run multiple development streams simultaneously.

Slash Commands and Radical Automation

At the heart of Cherny's productivity system are slash commands, custom shortcuts committed directly into the project repository. These one-keystroke macros automate complex multi-step operations that would normally require manual orchestration. His most-used command, /commit-push-pr, handles everything from staging changes to opening a pull request, and Cherny says he invokes it dozens of times each day.

The implications are staggering. What used to be a 10-minute ceremony of git operations, code review preparation, and CI pipeline monitoring is now a single keystroke followed by a few seconds of wait time. Multiply that across an entire engineering organization and the productivity gains become enormous.

The CLAUDE.md Knowledge File

Perhaps the most replicable element of Cherny's system is what he calls the CLAUDE.md file. This is a single markdown document checked into the project's git repository that serves as institutional memory for the AI agent. Every time Claude makes a mistake or misunderstands a pattern, the team documents it in CLAUDE.md so the same error never recurs.

Think of it as a living style guide crossed with a bug tracker, except it is written for an AI reader rather than a human one. Over time, this file becomes an increasingly precise specification of the project's conventions, edge cases, and architectural decisions. The result is an AI assistant that gets smarter with every commit.