newsfilter.io

Inside Claude Code With Its Creator Boris Cherny

Y CombinatorBoris Cherny, Mark Mandelmann, Jr., Francesc Campoy, Melanie Warrick
Watch on YouTube →

  • Core Strategic Philosophy

    • Anthropic builds for models six months in the future rather than optimizing for current capabilities, anticipating that "scaffolding" features will be obsolete once models improve.
    • Boris Cherny advises founders to identify the "frontier" where current models are weak, build around it, and wait for the model to naturally achieve the capability without manual engineering.
    • The team adheres to the "bitter lesson": they never bet against the model's general capability, treating any custom code scaffolding as temporary technical debt.

  • Product Evolution & Metrics

    • The entire Claude Code codebase is rewritten and replaced approximately every six months; there is virtually no code in the current product that existed six months prior.
    • Anthropic reports that productivity per engineer has grown by 150% since the release of Claude Code.
    • At Anthropic, 70% to 90% of all code is written by Claude Code, with some teams (including the creator) reaching 100% reliance.
    • The product transitioned from a CLI to a terminal interface by accident; the initial goal was a chat app, but tool-use capabilities revealed the terminal as the most efficient medium.
    • Claude Code was built entirely by Claude Code in roughly 10 days for the desktop application (Co-work), demonstrating the model's ability to build its own ecosystem.

  • Feature Development & Iteration

    • Plan Mode: A feature born from user demand to "plan without coding"; currently used in 80% of sessions by the creator, though it is predicted to have a limited lifespan as models become capable of one-shot execution.
    • ClaudeMD: A configuration file that stores context and rules; it is designed to be short (under 2,000 tokens) to prevent context bloat, with a recommendation to delete and restart if it grows too large.
    • Verbosity: The team iterated on terminal output verbosity based on user feedback; initially, they suppressed bash output to save tokens, but user revolt led to the reintroduction of detailed logs and a "verbose mode."
    • Multimodal & Logging: Recent updates include summarizing file reads and searches (e.g., "searched one pattern") only after the model became reliable enough to do so without misreading files.

  • Workflow & Human-Machine Interaction

    • Agent Topologies: Teams utilize "subagents" (recursive Claude Code instances) to run parallel research tasks; the creator adjusts the number of subagents (e.g., 3, 5, or 10) based on task difficulty.
    • Specialization vs. Generalization: Effective engineering teams at Anthropic are bimodal, consisting of extreme specialists (e.g., in JavaScript runtimes) and hyper-generalists who span product, design, and infra.
    • Hiring: The team evaluates candidates on "beginner mindset" and scientific thinking, specifically asking for examples of when they were wrong to test their ability to learn and adapt to shifting model capabilities.
    • Evaluation: The company now considers uploading Claude Code transcripts as a valid metric for evaluating a candidate's system thinking, testing, and product sense.

  • Future Trends & Predictions

    • End of Traditional IDEs: The creator predicts the title "software engineer" will fade, replaced by "builder" or "product manager," as coding becomes a general skill accessible to designers, PMs, and non-technical staff.
    • UI Evolution: While the terminal remains central due to its "latent demand" and speed, Claude Code is expanding into web, mobile, and IDE extensions; the creator admits the terminal's dominance is an experiment rather than a permanent design constraint.
    • Safety & ASL: Anthropic tracks "AI Safety Levels," currently at ASL 3 (current models); ASL 4 involves recursive self-improvement, which presents both massive capability gains and catastrophic misuse risks (e.g., bioweapons, zero-days).
    • Market Adoption: Claude Code is reportedly the model of choice for 70% of startups, and accounts for 4% of all public code commits across the industry.

  • Co-work & Non-Technical Users

    • Co-work was developed to serve non-technical users by wrapping Claude Code in a GUI with sandboxed VMs, permission prompts, and safety guardrails.
    • The product expanded beyond coding to tasks like monitoring tomato plants and recovering corrupted photos, validating the "latent demand" for general automation tools.
    • Co-work allows internal Claude Code agents to communicate with Slack users and other agents to clarify requirements, automating the feedback loop.

  • Cultural & Operational Insights

    • Tool Use: The product strategy centers on enabling the model to use tools (APIs, file systems, bash) rather than constraining it; the model "wants" to interact with the world.
    • Rapid Prototyping: The CLI allows for 50-100 iterations of UX elements (like spinners) in hours, a speed impossible with traditional graphical prototyping tools.
    • Mission Alignment: The team operates with a heavy focus on AI safety, with safety discussions being more prominent in office culture than product features.