Anthropic Engineers Reveal How They Actually Use Claude to Write Code

In a detailed public post, Boris Cherny, the engineer behind Claude Code, has laid out exactly how his team at Anthropic uses their own AI tool to build software. The description shows a shift from treating AI as a simple helper to making it a core part of their engineering process.

The team’s central method is running multiple Claude sessions at once using a feature called git worktrees. Cherny says having three to five sessions open simultaneously is the single biggest boost to their output. Each session can focus on a different task, turning a linear process into a parallel one.

They also stress planning. For complex jobs, engineers start by having Claude draft a detailed plan. Sometimes, they even use a second Claude instance to review that plan like a senior engineer would, catching problems before any code is written. The rule is simple: if something starts going wrong, stop and re-plan.

Knowledge is managed actively. Engineers maintain a CLAUDE.md file, instructing the AI to update it after every mistake so it learns. This file is edited constantly, becoming a tailored guide for each project.

Repetitive tasks are automated into reusable ‘skills’ saved in their code repository. One engineer built a command that scans for duplicate code at the end of each session. Another created an agent that can write data models and test changes.

For debugging, the team gives Claude access to logs and Slack threads and lets it determine the fix. Cherny notes Claude often resolves bugs on its own when given proper context. The approach requires trust, delegating significant autonomy to the AI.

These practices, coming directly from the tool’s builders, suggest AI-assisted development is maturing into a disciplined engineering practice. It’s less about a magical autocomplete and more about redesigning the workflow around a capable, parallel partner.