Boris Cherny

Boris Cherny is the creator of Claude Code, Anthropic’s coding agent in the terminal. He developed its initial prototype and helped shape its approach to context, search and automation alongside a growing team. In May 2025, he was its lead engineer; Anthropic listed him as Head of Claude Code for its May 6, 2026 developer conference. His central design principle was to give users direct access to the model through a tool they could compose into larger workflows.[4][5][6]

Claude Code began with experiments Cherny conducted after joining Anthropic. Recalling that period in his May 2025 interview, he said, “There was no master plan.” He used the public API to build a terminal assistant that could react to his music listening and explain screenshots from a video player. Giving it terminal access and the ability to code made it useful enough for daily work. He described expanding access from the core team to Anthropic’s engineers and researchers, then deciding external users should be able to try it.[5]

Cat Wu named Cherny, Sid and Ben as the original team. She began sending feedback while using the tool for data analysis and visualizations; Cherny recalled asking that she become the project’s PM. Their work on the agent built on the underlying Claude model team’s contributions. Cherny later appeared with Claude Code Plan Mode among the launches covered in the June 2025 AI Engineer World’s Fair recap.[5][4]

Cherny distinguished three places to implement behavior: inside the model, in Claude Code’s scaffolding, or in an external workflow that uses Claude Code as a tool. Session management with Tmux, for example, could remain outside the agent. “We think of it as like a Unix utility,” he explained: something composable like grep or cat. This framing supported both interactive coding and non-interactive automation rather than requiring one prescribed development workflow.[5]

His May 2025 explanations made that preference concrete. The team tried truncating messages and rewriting old tool calls before choosing a simpler compaction mechanism: ask Claude to summarize previous messages. For user-directed memory, it shipped CLAUDE.md, a text file automatically read into context, with versions available in the repository root, child directories and the user’s home directory. These were practical mechanisms for supplying context without elaborate memory architectures.[5]

Cherny also described moving from early indexed retrieval-augmented generation experiments to agentic search: letting the model search code with ordinary tools such as glob and grep. He said the team preferred the results, while characterizing the comparison as mostly subjective internal experience with some unspecified internal benchmarks. The choice also avoided maintaining an index that could drift out of sync and storing sensitive code in an additional system. The tradeoff was greater latency and token use.[5]

One concrete workflow he described was the team’s semantic linter. A GitHub Action invoked Claude Code with a saved local slash-command prompt to check spelling, agreement between code and comments, and use of approved libraries. Claude made corrections and used the GitHub MCP server to commit them back to the pull request. The example illustrated his preference for composing prompts, automation and tools: checks difficult to express through static analysis could instead be written as Markdown instructions.[5]

In May 2025, Cherny estimated that roughly 80–90% of Claude Code’s code was Claude-written; Wu emphasized that it still received substantial human review. Cherny preferred doing some intricate data-model refactoring by hand because experimenting directly was easier than explaining his detailed preferences to the model. He also distinguished letting an agent iterate on tests from allowing potentially destructive shell commands. Wu credited Robert on the team with leading permission-system work.[5]

Cheaper implementation changed Cherny’s own design process: he said he could ask Claude Code for three prototypes and compare working alternatives instead of first writing a long design document. He recalled a particularly concrete example from the night before launch. After finding several existing Markdown libraries unsatisfactory for terminal rendering, he asked Claude to write a parser. He corrected the suggestion that it was zero-shot: it took roughly one or two prompts, and the resulting parser was the one used in Claude Code at the time of the interview.[5]

Cherny credited another Anthropic engineer, Jeremy, with changing his view of automated fixes. Jeremy built a bot that read a Slack feedback channel and used an early Claude Code version to propose fixes as pull requests. Cherny did not remember its success rate, but recalled that it fixed enough issues to make him a believer in the workflow.[5]

In his September 19, 2026 essay “I am often wrong”, Cherny described his approach to product work as an iterative sequence: understand the available information, gather what is missing, define the problem, choose a clear and simple approach, set a goal, and act urgently. New information could require revising the problem, approach and goal. “When there’s new data, you have to update your priors,” he wrote. He identified unclear problems and unnecessarily complex approaches as recurring failure modes, and said he expected reciprocal feedback, including challenges to the framework itself.[7]