Boris Cherny: How We Built Claude Code

Quick Overview

Boris Cherny, Creator & Head of Claude Code at Anthropic, advises founders building on LLMs not to build for the model of today but for the model six months from now, emphasizing that Anthropic's code, like Claude Code, has been constantly rewritten and that relying solely on external tools or highly specialized expertise can be limiting, favoring instead a generalist approach and continuous learning.

Key Points: Cherny advises founders building on LLMs to build for the model six months in the future, not the current one, as capabilities change rapidly. Anthropic's code base, including Claude Code, has been continuously rewritten and iterated upon over the last six months. He notes that many early AI applications relied heavily on the terminal (CLI) because building a graphical UI was too time-consuming. The development of QuantMD, an internal tool, was an accident driven by the need to debug memory leaks, leading to a useful product. He argues against over-specialization in AI development, stating that generalists who understand core principles and debugging (like reading logs) are often more effective than hyper-specialists in areas like formal verification. Anthropic's core principle is teaching the model how to use tools rather than hardcoding specific features, which allows the model to adapt better to future capabilities.

Context: Boris Cherny, Creator & Head of Claude Code at Anthropic, participates in a panel discussion with other Y Combinator partners (Garry Tan, Diana Hu, Jared Friedman) about building and scaling AI products, specifically focusing on the evolution of their internal coding tools and the philosophy behind developing for rapidly advancing LLM capabilities.

Detailed Analysis

Boris Cherny, Creator & Head of Claude Code at Anthropic, strongly advises founders building applications on large language models (LLMs) to design for the model six months out, not the current one, due to the rapid pace of AI evolution. He points out that Anthropic's internal code base, including Claude Code, has been constantly rewritten and refined over the past six months, illustrating the need for adaptability. Cherny recounts the accidental creation of QuantMD, an internal debugging tool born from the necessity to fix memory leaks, which proved highly useful for developers. He stresses that relying too heavily on hyper-specialized knowledge or overly complex tooling like formal verification can be limiting; he prefers generalists who understand core principles and debugging techniques. He contrasts this with the past, where early AI tools often relied on the terminal (CLI) because building a full GUI was too slow. A key principle at Anthropic is enabling the model to use tools effectively rather than hardcoding features, which allows the model to remain useful even as capabilities evolve. He notes that Anthropic's core product code is heavily typed (TypeScript), which helps prevent common errors, unlike older, less rigorous systems. He also highlights that their internal tools are designed to be generalist-friendly, contrasting with hyper-specialist tools, and that their productivity has increased significantly, with internal code reviews showing 70% of public commits coming from Anthropic engineers, a figure he finds surprisingly high.