How Claude Code’s Creator Actually Uses It

Boris Cherny — creator and head of Claude Code at Anthropic — has described a working style in which one person operates with the throughput of a small team. The trick isn’t typing faster. It’s a shift in role: from writing code by hand to directing a fleet of autonomous agents. Five principles carry the whole approach.

Run agents in parallel, not code in sequence

The core move is abandoning the linear write-test-fix loop. Instead of one task at a time, Cherny keeps several Claude instances running side by side, each on a distinct job — one running a test suite, another refactoring a module, a third drafting docs. System notifications tell him which agent needs input.

The developer becomes a commander assigning missions and managing execution, not a typist. Productivity scales by orchestration — running today’s models well in parallel — rather than by waiting for a faster model.

Use the smartest model to shrink the correction tax

Counterintuitively, Cherny works almost entirely with Opus, Anthropic’s heaviest and slowest model. Faster models like Sonnet or Haiku generate quicker, but they need more human guidance and correction.

The real bottleneck in AI-assisted development isn’t token-generation speed — it’s the human time spent fixing the agent’s mistakes. A stronger model raises the compute cost per task but sharply lowers that correction tax, so complex work reaches “done” sooner overall.

Keep a living rulebook in CLAUDE.md

Agents forget everything between sessions. The fix is a single CLAUDE.md file, committed to the repo, that acts as persistent instructions for the agent.

Whenever an agent makes a mistake or breaks a team convention, the developer adds a corrective rule to CLAUDE.md. Every human correction becomes a permanent lesson — the agent grows steadily more aligned with the project’s specific standards instead of relearning them each time.

Automate the bureaucracy

Repetitive, low-judgment work is pushed onto slash commands and specialized sub-agents:

• Slash commands — custom shortcuts checked into the repo that collapse multi-step operations into one call, e.g. a single /commit-push-pr that commits, pushes, and opens a pull request.
• Sub-agents — focused personas for specific phases, such as a code-simplifier for refactoring or a verify-app agent for end-to-end testing.

Make the agent prove its work

The most important habit is the verification loop. The agent isn’t just asked to write code — it’s required to show the code works.

Cherny gives his agents the means to test their own changes: running test suites, executing bash commands, even driving a web UI through a browser extension. That feedback forces the agent to iterate and self-correct until the output is functional, not merely plausible. Closing this loop is the single biggest lever on the quality of agent-written code.

Where to go next

This workflow has been packaged into an open-source plugin, Oh My Claude Code (OMC), which automates the same parallel, multi-agent orchestration with defined execution modes and model routing.

• Oh My Claude Code (OMC): Agent Swarm Orchestration — the productized version of this workflow.
• Claude Code Operator’s Guide — the configuration and multi-agent patterns behind it.