Detailed Analysis
Boris Cherny, the creator of Claude Code and Head of the product at Anthropic, has published a detailed account of his personal daily workflow using the tool he built, offering rare insider visibility into how the engineers closest to the technology actually deploy it. His setup, which he describes as "surprisingly vanilla," centers on running five simultaneous terminal instances of Claude Code across separate git checkouts of the same repository, complemented by five to ten additional web-based sessions on claude.ai/code, and even mobile handoffs via the Claude iOS app. He exclusively uses Opus 4.5 with thinking mode enabled, arguing that despite the model's greater size and latency compared to Sonnet, its superior tool use and reduced need for steering make it faster in practice — a counterintuitive but significant operational observation about the economics of large model deployment.
Several of Cherny's techniques reflect a philosophy of building systems and infrastructure around Claude Code rather than relying on one-off prompting. His team maintains a shared CLAUDE.md file checked into git, collaboratively updated multiple times per week with observed failure modes, effectively functioning as a living institutional memory for the AI. This extends to code review workflows, where tagging @.claude on pull requests triggers automatic updates to CLAUDE.md as part of the PR itself — a practice Cherny frames as "Compounding Engineering," borrowing a concept from writer Dan Shipper. Complementing this are shared slash commands stored in .claude/commands/, PostToolUse hooks that auto-format code to prevent CI failures, and granular permission pre-approvals via .claude/settings.json, all of which reduce friction and eliminate repetitive interruptions during long autonomous runs.
The workflow reflects a broader architectural principle that Cherny consistently articulates: AI coding agents produce their best results when given a reliable feedback loop for self-verification. His top tip — ensuring Claude can verify its own work — is operationalized through domain-specific methods, including using the Claude Chrome extension to visually test UI changes in a browser and iterate until correct. For other domains, verification takes the form of background agent checks, deterministic Stop hooks, or community-developed plugins. This emphasis on closing the feedback loop distinguishes Cherny's approach from simpler prompt-and-review workflows, positioning Claude Code as an autonomous system that continuously self-corrects rather than a tool that merely generates code for human inspection.
The publication of these tips carries significance beyond practical advice, as it reveals the internal operational norms developing at Anthropic around agentic AI use. The parallel-instance model — treating Claude Code sessions as distributed compute resources rather than single conversational threads — represents a meaningful shift in how software engineering itself is being reconceptualized. The use of subagents for routine PR workflows (build validation, code simplification, end-to-end verification) and integration with external services like Slack via MCP servers and BigQuery via CLI suggests a trajectory toward AI systems that are embedded deeply in engineering infrastructure rather than used as standalone assistants. The fact that Cherny's approach has already contributed to outcomes like COBOL-to-cloud migrations being compressed from months to weeks underscores the scale at which these practices are being applied.
Situating Cherny's workflow within broader AI development trends, his methods exemplify the emerging paradigm of "orchestration over conversation" — where the primary skill for advanced AI users is not prompting but designing the environment in which the AI operates. The emphasis on pre-approved permissions, shared configuration files, automated hooks, and parallel execution mirrors patterns seen in distributed systems engineering, suggesting that AI tooling at the frontier is converging with DevOps and infrastructure-as-code practices. As Claude Code's daily active users have reportedly doubled recently and the tool is increasingly adopted by non-technical users for tasks like legacy system migrations, Cherny's documentation of these practices serves both as a benchmark for professional users and as a signal of where the industry is heading: toward AI agents that run autonomously for extended periods, embedded in team workflows as first-class engineering collaborators.