# Boris Cherny: Building Claude Code Source: https://www.youtube.com/watch?v=qyPCVqFUyDo Recap page: https://rapidrecap.app/video/qyPCVqFUyDo Generated: 2026-07-28T00:25:11.221+00:00 --- ## The Gist Boris Cherny, head of Claude Code at Anthropic, reveals that Claude Code deletes 80% of its system prompts and requires developers to stop over-specifying and instead let frontier models like Opus 5 solve hard, ambiguous tasks autonomously using dynamic workflows and automated verification. ## Quick Overview Boris Cherny outlines how Claude Code achieves massive autonomy through radical prompt reduction, unhobbling frontier models, and running thousands of autonomous agents in parallel. Rather than micro-managing models with complex step-by-step instructions, developers must provide clear guardrails, exit criteria, and automated verification loops, allowing models to operate like experienced software engineers. **Key Points:** - Anthropic launched Opus 5 with a 30 percent performance improvement over Arc-AGI 3. - Claude Code deleted over 80 percent of its system prompts by trusting the raw intelligence of newer model generations. - The Claude Code desktop app was successfully rewritten, tested, and verified in Swift using autonomous agent workflows over a two-week period. - A single Claude Code dynamic workflow can orchestrate thousands of parallel agents to handle complex code refactoring, bug fixes, and data analysis. - Traditional prompt engineering is being replaced by empirical testing, where developers provide hard tasks and verification loops rather than rigid instructions. - Boris Cherny advises developers to avoid listening to LinkedIn influencers and instead learn computer science through practical, problem-solving code implementation. ![Screenshot at 25:02: Boris Cherny and Diana Hu discussing the massive scale of autonomous agent workflows running across thousands of instances.](https://ss.rapidrecap.app/screens/qyPCVqFUyDo/00-25-02.jpg) **Context:** At Y Combinator Startup School 2026, Managing Partner Diana Hu interviews Boris Cherny, Head of Claude Code at Anthropic, to discuss the practical realities of building agentic software engineering products on top of rapidly accelerating AI models like Opus 5. ## Detailed Analysis Boris Cherny explains that the defining shift in modern AI development is moving away from rigid prompt engineering and micro-management toward empirical science. By deleting 80 percent of system prompts and removing legacy scaffolding, Anthropic allows Opus 5 to operate with minimal friction. Developers should stop over-specifying tasks and instead give models hard problems, strict exit criteria, and automated verification tools. Through dynamic workflows, loops, and routines, a single developer can orchestrate thousands of agents to handle complex multi-file codebases, write tests, and maintain production software autonomously. ### What Makes Opus 5 Different - Opus 5 delivers a massive leap in model intelligence and capability, pushing ARC-AGI 3 scores up by 30 percent compared to single-digit scores in older versions. - The model runs for significantly longer periods without stopping, allowing it to execute extended reasoning and multi-step problem solving. - Opus 5 is no longer prompt-injectable in the way older models were, eliminating the lethal trifecta vulnerability that allowed malicious internet text to override user computers. ![Screenshot at 02:09: Diana Hu and Boris Cherny discussing the performance leaps in Opus 5.](https://ss.rapidrecap.app/screens/qyPCVqFUyDo/00-02-09.jpg) ### Press Delete on Your AI Product - Claude Code deleted over 80 percent of its system prompts because newer models possess innate capabilities that no longer require hand-holding. - Every time a new model generation is released, developers should perform an ablation study, stripping away old system prompts, tools, and hooks line by line. - Attempting to carry over legacy scaffolding and prompt instructions from past models creates product overhang that actively gets in the model's way. ![Screenshot at 05:51: Boris Cherny explaining the importance of deleting legacy system prompts.](https://ss.rapidrecap.app/screens/qyPCVqFUyDo/00-05-51.jpg) ### Product Overhang and Unhobbling AI - Product overhang occurs when today's models are capable of complex behaviors that developers have not yet realized or enabled. - Anthropic discovered that Opus 5 can natively leverage external tools like OpenCV to draw landscapes, portraits, and animals without explicit drawing training. - Developers must embrace an empirical mindset, experimenting freely with models to uncover hidden capabilities rather than assuming strict operational boundaries. ![Screenshot at 12:19: Boris Cherny detailing the concept of product overhang.](https://ss.rapidrecap.app/screens/qyPCVqFUyDo/00-12-19.jpg) ### Prompt Engineering Is Changing - The most common engineering mistake is over-specifying instructions and forcing models to follow rigid, step-by-step sequential workflows. - Modern prompting requires defining high-level tasks, guardrails, and exit criteria, then letting the model figure out execution independently. - Verification is the single most critical component, requiring developers to provide automated testing tools so models can check their own work and self-correct. ![Screenshot at 21:55: Boris Cherny discussing the shift away from rigid prompt engineering.](https://ss.rapidrecap.app/screens/qyPCVqFUyDo/00-21-55.jpg) ### Running Thousands of AI Agents - Dynamic workflows allow developers to orchestrate dozens or thousands of parallel AI agents to tackle massive code refactoring projects. - Routine automation handles daily tasks such as dead-code cleanup, test coverage generation, and abstraction policing without human intervention. - Functional programming principles applied to agent architectures enable clean sequencing and parallel execution across complex codebases. ![Screenshot at 25:35: Boris Cherny explaining how dynamic workflows orchestrate thousands of agents.](https://ss.rapidrecap.app/screens/qyPCVqFUyDo/00-25-35.jpg) ### What Every CS Student Should Still Learn - Coding is solved for routine tasks, but exceptional builders differentiate themselves through strong design sense, business intuition, and user empathy. - Learning computer science practically by building real tools and solving hard personal problems remains essential for developing deep technical intuition. - Students should focus on mastering core fundamentals and learning how to apply technology rather than chasing viral advice from influencers. ![Screenshot at 32:38: Boris Cherny offering advice to computer science students and founders.](https://ss.rapidrecap.app/screens/qyPCVqFUyDo/00-32-38.jpg)