Maximizing Productivity with Claude Code: Insights from Boris Cherny of Anthropic

Boris Cherny, a key figure at Anthropic, has a wealth of knowledge about enhancing productivity using Claude Code. He shares some compelling practices that not only streamline development but also improve collaboration and learning within teams. This article delves into his methodologies, operational strategies, and the tools he uses to maximize the output of Claude Code.

Leveraging Parallel Instances for Enhanced Efficiency

Cherny runs multiple instances of Claude Code in parallel to boost productivity. Using his MacBook’s terminal, he executes five local sessions and another 5-10 sessions through Anthropic’s website. To maintain performance and avoid conflicts, each local session operates with its own git checkout rather than branches or worktrees. This approach allows him to work simultaneously across various projects without interruptions.

Additionally, he employs command-line instructions like & to initiate remote sessions, often utilizing --teleport for swift transitions between tasks. Despite the streamlined process, he acknowledges that around 10-20% of sessions may be abandoned due to unforeseen circumstances, which is something to keep in mind in this fast-paced environment.

Choosing the Right Tools: Opus 4.5 vs. Sonnet

Cherny has a preference for using Opus 4.5 over Sonnet for coding tasks. While Sonnet may be faster, he considers Opus to deliver better quality and reliability. Cherny emphasizes that Opus excels in tool usage and, ultimately, proves to be faster overall when factoring in the smaller model’s performance.

Documentation and Continuous Learning with CLAUDE.md

At Anthropic, learning from past experiences is crucial. Each team maintains a CLAUDE.md file in git dedicated to documenting mistakes and best practices. This includes style conventions, design guidelines, and PR templates. Cherny uses the @.claude tag on his coworkers’ PRs to enrich the CLAUDE.md file, ensuring that knowledge is preserved and accessible for future reference. Currently, their CLAUDE.md holds 2.5k tokens, showcasing a wealth of information that benefits the team.

More Read

Key Insights from Speech Reasoning Language Models: Valuable Lessons Learned

Optimizing Educational Assignment Feedback: A Comprehensive Framework Using LLM Agents for Synthetic Generation

Understanding Effective Captions: A Detailed Benchmark for Evaluating Visual Caption Quality in Correctness and Thoroughness

RogueMerge: Comprehensive and Unified Strategies for Attacking LLM Model Merging

Optimizing Multi-Task Speech Models: Efficient Distillation with Language-Specific Experts

Structured Workflows: Planning and Auto-Editing

One hallmark of Cherny’s workflow is starting with a clear plan. He believes in refining this plan collaboratively with Claude before transitioning to an auto-editing phase. He notes:

If my goal is to write a Pull Request, I will use Plan mode, and go back and forth with Claude until I like its plan. From there, I switch into auto-accept edits mode and Claude can usually 1-shot it. A good plan is really important!

Streamlined Daily Workflows with Slash Commands

Cherny employs slash commands to execute daily tasks, such as commits, PRs, and verifications, by using sub-agents. All commands are neatly organized in .claude/commands/, which not only saves time but minimizes the need for explicit prompts. For instance, the /commit-push-pr command combines inline bash operations to pre-compute necessary information, enabling quick execution and reducing back-and-forth interaction with Claude.

Ensuring Code Quality with PostToolUse Hooks

Despite the generally well-formatted code generated by Claude, inconsistencies can lead to CI failures. Cherny combats this issue by implementing a PostToolUse hook:

"PostToolUse" : [
    "matcher": "WritelEdit",
    "hooks": [
        {
            "type": "command",
            "command": "bun run format || true"
        }
    ]
]

Security Measures in Workflows

Security is a top priority for Cherny, who almost never utilizes the --dangerously-skip-permissions option. Instead, he sets up safe, commonly used bash commands through /permissions to prevent unnecessary permission prompts for tasks like bun run build:* and bun run test:*. The only time he resorts to the risky option is for long-running tasks within a sandbox environment, which mitigates the risk of interruptions.

Creating Feedback Loops for Quality Assurance

A crucial component of Cherny’s workflow is establishing feedback loops, allowing Claude to verify its outputs by executing bash commands, running test suites, or simulating app usage. He asserts that such feedback mechanisms can enhance the final results significantly:

Claude tests every single change I land to claude.ai/code using the Claude Chrome extension. It opens a browser, tests the UI, and iterates until the code works and the UX feels good.

Collaborative Code Reviews and Team Focus

Ultimately, Cherny’s methodologies enable his team to concentrate on critical tasks like code review and steering. By implementing these structured workflows, the code presented for review is typically polished and well-prepared, allowing engineers to focus on higher-level evaluations rather than getting bogged down by formatting issues.

Recent tweets from Cherny have sparked lively discussions on X.com regarding these practices, offering even more insights. For anyone interested in elevating their development processes, exploring Cherny’s techniques with Claude Code provides a promising roadmap to enhanced productivity.

Inspired by: Source