When Boris Cherny, the creator of Claude Code—the most advanced AI coding agent—shares insights, the tech world listens closely. Over the past week, Cherny’s detailed explanation of his personal development setup has captured the attention of Silicon Valley and developers worldwide, evolving into a viral discussion about the future of software engineering.
Cherny’s thread on X (formerly Twitter) has been hailed by industry figures as a pivotal moment. Jeff Tang, a respected developer community leader, emphasized that programmers not familiar with Cherny’s best practices are falling behind. Meanwhile, Kyle McNease suggested Anthropic, Cherny’s company, might be experiencing its own breakthrough akin to ChatGPT’s rise.
Orchestrating Multiple AI Agents: Coding as a Real-Time Strategy Game
The hallmark of Cherny’s approach is running five Claude AI agents concurrently within his terminal environment. Unlike traditional sequential coding—writing, testing, and moving on—Cherny acts as a commander overseeing multiple streams of work simultaneously. He assigns each Claude instance a numbered tab and uses iTerm2 system notifications to manage inputs and responses efficiently.
Each AI agent handles distinct tasks: one runs test suites, another refactors legacy code, and yet another drafts documentation. Additionally, Cherny operates several Claude sessions in a web browser, using a specialized “teleport” command to seamlessly transfer sessions between his local machine and the cloud.
This multi-agent orchestration exemplifies Anthropic’s “do more with less” philosophy, championed by President Daniela Amodei. While competitors invest heavily in infrastructure, Anthropic focuses on optimizing model coordination to exponentially boost developer productivity.
The Strategic Choice of the Slowest, Smartest AI Model
Contrary to the industry’s typical race for speed, Cherny exclusively employs Anthropic’s largest and slowest model, Opus 4.5. He explains that despite its slower token generation, Opus 4.5 requires less human correction and demonstrates superior tool use, ultimately accelerating overall development.
This insight is critical for enterprise leaders: the true bottleneck isn’t AI speed but human time spent fixing AI errors. Investing in a smarter, more deliberate model reduces the cumulative “correction tax,” enhancing efficiency.
Transforming Mistakes into Learning with a Shared Knowledge Base
Addressing the common challenge of AI “amnesia,” Cherny’s team maintains a single file named CLAUDE.md within their git repository. Whenever the AI makes an error, the team documents the mistake and corrective instruction in this file, ensuring that Claude progressively learns and adapts over time.
This process effectively turns the codebase into a self-improving system. Human reviewers don’t just fix bugs; they update the AI’s behavioral rules, making every error a permanent lesson. Product leader Aakash Gupta notes that this continuous learning loop significantly increases the agent’s intelligence and reliability.
Automating Tedious Tasks with Slash Commands and Specialized Subagents
Cherny’s workflow heavily relies on automation to handle repetitive development tasks efficiently. He utilizes custom slash commands embedded in the project repository to trigger complex operations with minimal input.
One prominent example is the /commit-push-pr command, which automates the entire version control process—including committing changes, pushing to the repository, and opening pull requests—saving time and reducing manual errors.
Additionally, Cherny deploys AI subagents tailored for specific responsibilities, such as simplifying code architecture post-development and executing comprehensive end-to-end application tests before deployment.
Verification Loops: The Key to High-Quality AI-Generated Code
A major factor behind Claude Code’s reported achievement of $1 billion in annual recurring revenue is its robust verification loop. Unlike basic text generation models, Claude actively tests every code change it makes using browser automation and other tools.
Cherny explains that the AI opens the browser, interacts with the user interface, runs tests, and iterates until the code functions correctly and offers a good user experience. This self-verification improves output quality by two to three times, ensuring the agent doesn’t just write code, but confirms its correctness.
Implications for the Future of Software Engineering
The enthusiastic response to Cherny’s workflow signals a transformative shift in software development paradigms. Whereas AI-assisted coding was once limited to autocomplete features, Claude Code demonstrates the potential for AI to act as a full-fledged workforce, dramatically multiplying human productivity.
Jeff Tang summarizes the opportunity for engineers succinctly: embracing these AI systems offers unprecedented power. Developers willing to rethink AI as a team member rather than merely an assistant will gain a decisive advantage, effectively playing an entirely new game while others continue typing.
Related Articles
YouTube Tests New Feature to Enhance Home Feed Customization
U.S. Energy Department Shifts Focus from Renewables to Fusion Amid Organizational Overhaul
Siemens Launches AI-Powered Agent to Revolutionize Automation Engineering
OpenAI’s Sweetpea AI Wearable Aims to Rival Apple’s AirPods with Advanced Features
