Skip to content

Boris Cherny: Building Claude Code

Y Combinator Podcast

36m
Transcript

Original podcast

Where to listen

Summary

Written by a model from the transcript; not yet reviewed.

This episode covers the creation and capabilities of Claude Code and the newly shipped Opus 5 model, featuring a discussion on how modern AI agents can run for extended periods, rewrite large codebases, and maintain software autonomously. The guest and host explain why traditional development practices like over-specifying system prompts and rigid system design are obsolete, arguing instead for an empirical approach where developers "unhobble" models by deleting system prompts and letting agents verify their own work. Listeners will gain practical insights into leveraging dynamic workflows, test-time compute, and automated routines that spawn thousands of agents to handle tasks ranging from dead code removal to complete language migrations.

  • Delete the vast majority of your system prompts and custom rules when working with new model generations, as modern models like Opus 5 often perform better without them. 7:34
  • Give the model challenging, high-level tasks with clear guardrails and verification methods rather than overly specific, step-by-step instructions. 15:39
  • Utilize dynamic workflows and local or cloud-based loops to let Claude orchestrate thousands of agents for complex, multi-stage engineering tasks. 25:51
  • Build practical software and focus on real-world problem-solving, product sense, and user interaction alongside computer science fundamentals. 34:34

Chapters

  1. 0:07
    Podcast Intro and Opus 5 Launch

    Hosts welcome Boris Cherny and discuss the recent release of Opus 5 and its performance gains.

  2. 1:34
    Opus 5’s New Capabilities

    Boris explains Opus 5’s ability to run long tasks, avoid prompt injection, and its improved alignment.

  3. 3:53
    System Prompt Evolution in Claude Code

    Discussion of how Claude Code deletes and rebuilds its system prompt with each model release to match new behavior.

  4. 8:01
    Building and Rebuilding the Harness

    Boris outlines the iterative ablation process of removing and re‑adding prompt lines to find the minimal effective harness.

  5. 9:18
    Coding is Solved? The Future of Software Development

    Discussion on whether coding is solved, what remains challenging, and the skills that differentiate top builders.

  6. 11:09
    Unhobbling Models and Product Overhang

    The concept of unhobbling Claude and exploiting product overhang to unlock hidden model capabilities is introduced.

  7. 16:07
    Practical Examples of Model Overhang

    Boris shares real‑world use cases, such as codebase rewriting and creative tasks like drawing with OpenCV.

  8. 25:44
    Dynamic Workflows and Agent Orchestration

    Explanation of dynamic workflows, loops, and routines that launch thousands of agents for complex tasks.

  9. 28:58
    Claude’s Self‑Maintenance Routines

    Claude runs daily routines to clean code, write tests, and unify abstractions across multiple apps.

  10. 33:00
    Learning to Build with AI: Practical Advice for Students

    Boris shares his personal learning journey and advises students to focus on practical, product‑oriented skills.

  11. 35:12
    Closing Remarks and Special Offer

    Hosts wrap up, announce a special offer for attendees, and encourage building long‑running agentic projects.

Highlights

  • 1:30

    Opus 5 Long-Horizon Autonomyinsight

    Opus 5 introduces the capability to run autonomously for days, weeks, or even months at a time without requiring external scaffolding.

  • 2:32

    Eliminating Prompt Injectiondata

    By combining three years of alignment research, a traffic prompt injection classifier based on mechanistic interpretability, and an auto mode classifier, Opus 5 has effectively neutralized prompt injection attacks.

  • 3:53

    Deleting 80 Percent of System Promptsstrategy

    The engineering team deleted over 80 percent of the system prompt for Claude Code because the new model is intelligent enough to execute tasks without legacy correction prompts.

  • 9:58

    Constant Evals Over Static Codebasesinsight

    While codebases and system prompts should be routinely deleted and rebuilt with every new model generation, evaluation sets are kept and appended to until they are completely maxed out.

  • 17:46

    Rewriting Bun from Zig to Ruststory

    Using a dynamic workflow with Claude Code, an engineer had the model rewrite the entire Bun JavaScript runtime codebase from Zig to Rust over an 11-day period.

  • 21:55

    Two-Week Autonomous GUI Rewritestory

    The guest launched an ongoing experiment where Claude runs in a Mac virtual machine, rewrites the Electron desktop app in Swift, compares screenshots pixel-by-pixel, and updates a Slack channel every few minutes.

  • 28:33

    Automating Codebase Maintenance Routinesstrategy

    Claude maintains its own codebases across CLI, iOS, Android, and desktop apps by running daily routines that clean up dead code, unify duplicate abstractions, and ship experiments.

Quotes

  • “It has these skills, it has abilities that you actually didn't really teach it, but it just kind of learned.”
    Unnamed speaker · 1:19
  • “The model does not seem to be prompt injectable anymore.”
    Unnamed speaker · 2:02
  • “We deleted 80% of the system prompt.”
    Unnamed speaker · 4:51
  • “And the verification, I think, is probably the single most important thing that people do not get right virtually.”
    Unnamed speaker · 20:51
  • Max 20X 35:00–35:07

About this episode

Fresh off the launch of Opus 5, Claude Code creator Boris Cherny joins Diana Hu at Startup School 2026 to talk about what the newest models can do, how Claude Code came to be, and what it means to build products when the underlying capabilities keep accelerating.

Transcript

0:00
36:13

6,660 words · 2 speakers

Topics

  • Claude Code
  • Opus 5
  • Bun
  • routines
  • test time compute
  • Zig
  • dynamic workflows
  • Electron
  • Slack
  • OpenCV

Page updated · transcript and summary are machine-generated