How Claude Code is built (newsletter.pragmaticengineer.com)

Anthropic’s Claude Code — a terminal-first dev tool that has surged to over $500M annual run-rate and 10x usage growth since going GA in May — is built as a thin client around the Claude model rather than a traditional, logic-heavy application. Originating from a small prototype that let Claude interact with the filesystem, the product grew rapidly inside Anthropic, reaching heavy internal adoption (80%+ of engineers) and boosting PR throughput by 67% even as headcount doubled. The founding team reports extreme iteration velocity (roughly five releases per engineer per day, 10+ prototypes per feature) and rapid feature builds like “subagents” (a three-day build). They also emphasize product-overhang: the model already had capabilities that the product needed to unlock. Technically, Claude Code uses an “on-distribution” stack the model already understands: TypeScript, React with Ink for terminal UIs, Yoga for layout, Bun for build/packaging, and npm distribution. That lets Claude write most of its own code (≈90%), define UIs, call tools, traverse codebases and run tests while the client keeps minimal business logic. The result showcases an “AI-first” engineering workflow — agents for code review, TDD revival, automated incident response, and a lightweight shell that surfaces model capability with less scaffolding. Anthropic frames this as a testbed for model safety and product learning, though they note aspects of their approach (hiring, safety focus) may not transfer directly to all orgs.