Boris Cherny: Building Claude Code
Boris Cherny discusses Claude Opus 5's capabilities and the philosophy behind building Claude Code, emphasizing that modern AI models require a fundamentally different approach to product development. Rather than traditional engineering practices, builders should use empirical testing, aggressive prompt deletion, and continuous ablation to unlock each model's true capabilities.
Summary
Boris Cherny, creator of Claude Code, explains how Opus 5 represents a major leap forward from previous Claude models. Key capabilities include the ability to run for extended periods (days, weeks, months) in Auto Mode without stopping, and apparent resistance to prompt injection attacks through a combination of alignment work and mechanistic interpretability-based classifiers that can detect when prompt injection neurons activate in the model's weights.
Cherny reveals that Claude Code deleted 80% of its system prompt for Opus 5, a practice he applies with every new model release. Rather than building permanent systems, the team uses ablation methodology—deleting the entire system prompt and bringing it back line by line to understand the impact of each component. This approach stems from the reality that each model generation has different capabilities and "personality," requiring teams to relearn what instructions are actually necessary.
The concept of "product overhang" emerges as central to Cherny's philosophy. This refers to capabilities models possess but products fail to elicit due to unnecessary constraints and scaffolding. The original Claude Code exemplified fixing this overhang when Sonnet 3.5 could write entire files and features but existing products only offered single-line autocomplete. Cherny argues there are currently hundreds of unrealized product overhang opportunities in modern models.
Cherny demonstrates model capabilities through concrete examples: Opus 5 can rewrite entire codebases between programming languages (the Bun team rewrote 100,000+ lines from Zig to Rust in 11 days), and can perform unexpected tasks like creating portraits and landscapes using OpenCV without explicit training. He describes a personal experiment where he prompted Claude to rewrite an Electron app in Swift while comparing pixel-by-pixel to screenshots—a task that has been running for over two weeks with thousands of spawned agents.
The discussion covers dynamic workflows as a new form of test-time compute orchestration, where multiple agents run in sequence and parallel across stages, enabling productive execution of massive engineering tasks. Routines and loops allow continuous background maintenance tasks, with examples including automated dead code cleanup, test coverage analysis, and "abstraction police" that identifies and unifies nearly-duplicate abstractions across codebases.
Cherny addresses the future of software engineering now that coding is largely solved, arguing the key differentiator is empirical mindset rather than theoretical knowledge. He emphasizes forgetting prior assumptions about AI capabilities and being willing to try approaches that failed with previous models. His advice to students is to learn computer science through practical problem-solving and to combine technical skills with design sense, business understanding, and user empathy.
About this episode
<p>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.</p>
Key Insights
- Opus 5 can run continuously for days, weeks, or months in Auto Mode without traditional scaffolding like slash commands, representing a significant shift from previous models that required more explicit guidance.
- Claude Code deleted 80% of its system prompt for Opus 5 because the model's increased intelligence made many correction instructions unnecessary, establishing a pattern of aggressive deletion with each new model release.
- Product overhang—the gap between what models can do and what products let them do—represents a major opportunity for builders, exemplified by how the original Claude Code removed autocomplete constraints to let Sonnet 3.5 write entire files.
- The Bun team successfully used Opus 5 to rewrite over 100,000 lines of JavaScript runtime code from Zig to Rust in 11 days with a simple prompt and steering, a task that would have taken multiple engineers over a year in traditional software development.
- Prompt injection resistance in Opus 5 comes from combining three layers: alignment work spanning three years of research, mechanistic interpretability-based classifiers that detect which neurons activate during injection attempts, and Auto Mode classifiers—a technical achievement no competitor has yet demonstrated.
- Dynamic workflows represent a new form of test-time compute that orchestrates thousands of agents across multiple stages using an algebra-based system supporting sequence and parallel execution, enabling productive solutions to complex problems.
- Models demonstrate unexpected capabilities through proper solicitation—Opus 5 can create quality portraits and landscapes using OpenCV without explicit training, discovered accidentally through creative experimentation rather than directed development.
- The skill separating exceptional builders from average ones in the AI era is empirical mindset and willingness to abandon prior assumptions, rather than theoretical computer science knowledge or traditional engineering practices.
Topics
Transcript
. All right, Boris, we're so excited to have you here, the creator of Plot Code. Thank you.. It's great to be here. Fresh off the press, you guys just shipped Opus 5 yesterday. Yes. And it seems that model performance keeps accelerating. You guys took Arc AGI 3 to 30 percent, which is incredible. Yes. And for context, before the best score was in the low single digits or low teens, right? What can Opus 5 do now that it couldn't versus the previous version? Yeah, there's a lot that goes into every new model. And there's a lot of new capabilities that we teach and get the model to do. Whenever you do model training, you try to…
Full transcript available for MurmurCast members
Sign Up to AccessMore from Y Combinator Startup Podcast
Alexandr Wang: “This is a Once-in-a-Civilization Opportunity”
Alexandr Wang discusses his journey from math olympiad competitor to founding Scale AI, emphasizing the importance of developing independent conviction about the future while working on AI as a once-in-a-civilization opportunity. He shares insights on building Meta's AI research lab, the shift from human intelligence as a bottleneck to vision and ambition as the scarce resource, and practical approaches to agentic systems.
Blake Scholl: Breaking the Supersonic Ban
Blake Scholl, founder of Boom Supersonic, discusses building the first independently developed supersonic jet (XB-1), breaking regulatory barriers to legalize supersonic flight in the US, and how applying software development principles to hardware manufacturing can reduce costs and accelerate iteration in aerospace. He emphasizes the importance of working on problems you love, building in public to inspire others, and challenging the assumption that good ideas always get built quickly.
Sam Altman: "Never a Better Time to Do a Startup"
Sam Altman discusses why this is an unprecedented time to start companies, reflecting on Y Combinator's evolution, OpenAI's journey, and how AI is fundamentally changing what's possible for entrepreneurs. He emphasizes the importance of conviction in contrarian ideas, building strong networks, and maintaining earnestness in startup culture.
World Models, Explained
The episode explains world models—systems that predict future states in response to actions—as a critical path toward AGI and sample efficiency in AI. It contrasts model-free approaches (like language models) with model-based approaches using examples from chess, Go, self-driving cars, and robotics, showing how world models enable training on synthetic data and test-time planning.
How To Better Understand Your Users
The speaker advocates for using dot plots—a two-dimensional visualization showing individual user behavior over time—to understand how users actually interact with products, rather than relying solely on aggregate metrics like DAUs. Dot plots reveal usage patterns, feature adoption, and retention issues that aggregate data masks, and can scale from small user bases to millions of users through sampling.