Anthropic's Boris Cherny on Building Claude Code: Unlocking AI Potential — anthropic boris cherny building claude code
In a recent discussion at Y Combinator's Startup School, Boris Cherny, Head of Claude Code at Anthropic, provided deep insights into the development of their advanced AI model, Opus 5, and the philosophy driving Claude Code's evolution. Cherny's discussion emphasized a forward-thinking approach to AI development, focusing on 'unhobbling' models to achieve their full potential.
Opus 5: A Leap Forward in AI Performance
Cherny announced the recent shipment of Opus 5, Anthropic's latest AI model. This new iteration boasts significantly accelerated performance, a key indicator of its advanced capabilities. A major milestone highlighted was Opus 5 achieving a 30% score on the Arc AGI benchmark. This represents a substantial leap from previous scores, which were typically in the low single digits or teens, signaling a major advancement in AI model capabilities.
Furthermore, Opus 5 introduces an enhanced 'Auto Mode,' allowing the AI to operate for extended durations – potentially days, weeks, or even months – without requiring external scaffolding. This capability means the model can inherently understand its task and continue working towards completion autonomously.
Combating Prompt Injection and the 'Press Delete' Philosophy
Opus 5 also demonstrates remarkable resistance to prompt injection, a persistent challenge in AI development. Cherny revealed that this resilience, present since Opus 4.7, has been significantly enhanced in Opus 5. This is attributed to a combination of Anthropic's three years of research into AI alignment, a specialized prompt injection classifier that analyzes the model's neural activity, and an auto-mode classifier. These layers collectively make demonstrating prompt injection exceedingly difficult.
Anthropic employs a rigorous method for updating Claude Code, which Cherny described as a 'press delete' philosophy. With each new model release, they essentially reset the codebase and system prompt, starting from scratch. This process, known as 'ablation' in research, involves deleting the entire system prompt and then reintroducing it element by element to precisely understand the impact of each component. This meticulous approach allows them to adapt the framework to the unique characteristics of each new model. For instance, the advanced intelligence of Opus 5 allowed for the deletion of approximately 80% of the previous system prompt, as the model now intrinsically understands many behaviors that previously needed explicit instruction. Cherny encourages users to experiment with this by setting their own system prompts or utilizing the undocumented feature claude code simple=1, which removes all system prompts, including those for tools. He noted that models can sometimes perform better without extensive explicit prompts, although some are still necessary for a user-friendly product experience.
'Unhobbling Claude': Unleashing Latent AI Potential
A core concept introduced by Cherny is 'unhobbling Claude,' which involves removing product-imposed limitations that prevent AI models from realizing their full potential. He elaborated on the idea of 'product overhang,' describing the phenomenon where current AI models possess untapped capabilities that existing products fail to elicit. This often occurs because the product's design inadvertently 'hobbles' the model, restricting its ability to express its full range of abilities. Cherny referenced the initial version of Claude Code, built on Sonnet 3.5, as an example. While it was a powerful coding model, existing products were limited to single-line autocompletes or read-only chat interactions. The vision for Claude Code was to create a framework that eliminated these constraints, enabling the model to write entire files and features, thereby accessing its 'product overhang.'
Cherny believes significant opportunities exist for startups to identify and leverage this product overhang with current AI models. He advises founders to challenge models with tasks that appear to be slightly beyond their perceived limits, clearly define guardrails and exit criteria, and then allow the AI to perform. He shared an anecdote about a user discovering Opus 5's ability to generate images using OpenCV, enabling it to create portraits and landscapes simply through appropriate prompting—a capability that emerged organically rather than through explicit training.
The Future of AI and Coding
Cherny posited that coding is becoming increasingly 'solved' for many tasks, though complex challenges remain in areas like highly intricate systems code and distributed systems. He raised the question of whether AI agents could eventually write 100% of code, noting that while the desktop application is now notably fast, the journey toward this goal is ongoing. For those studying computer science, especially before the advent of AI agent coding, Cherny recommended focusing on practical application through startups or personal projects, rather than solely on theoretical knowledge. He emphasized that combining computer science skills with business acumen, user understanding, and data science knowledge is where true value resides in the evolving AI landscape.
This discussion underscores the importance of an empirical and experimental approach to leveraging AI. By understanding the current model's behavior through experimentation and shedding preconceptions from older models, users can unlock the true potential of advanced AI systems like Claude Code. For a deeper dive into the future of AI and Claude's role, explore insights on the anthropic claude code creator future.
Top comments (1)
I was particularly intrigued by the 'press delete' philosophy Anthropic employs for updating Claude Code, where they reset the codebase and system prompt with each new model release. This approach, also known as ablation, allows for a meticulous understanding of the impact of each component and enables the framework to adapt to the unique characteristics of each new model. The example of Opus 5, where approximately 80% of the previous system prompt was deleted due to the model's intrinsic understanding of many behaviors, highlights the potential of this method. I wonder if this approach could be applied to other AI development workflows, and what potential trade-offs might arise from adopting such a rigorous updating methodology.