Anthropic's Boris Cherny on Building Claude Code

In a recent discussion at Y Combinator's Startup School, Boris Cherny, Head of Claude Code at Anthropic, shared insights into the development and capabilities of their latest AI model, Opus 5, and the innovative approach behind Claude Code.

Opus 5: Performance and New Capabilities

Cherny announced that Anthropic had just shipped Opus 5 the day prior, highlighting its accelerated model performance. He noted the model's achievement of a 30% score on the Arc AGI benchmark, a feat previously in the low single digits or teens. This significant jump signifies a major advancement in AI model capabilities.

The full discussion can be found on YC's YouTube channel.

One of the standout new capabilities of Opus 5 is its ability to run for extended periods, especially when combined with 'auto mode.' Cherny described it as being able to 'go for days, weeks, months at a time,' without stopping and without needing external scaffolding like `slgo`. The model inherently understands its task and continues until completion.

Another notable advancement is Opus 5's apparent resistance to prompt injection, a problem that has plagued AI models. Cherny explained that this feature has been present since Opus 4.7, but Opus 5 takes it to a new frontier. This is achieved through a combination of Anthropic's three years of research into alignment, a prompt injection classifier that analyzes neural activity in the model's 'brain,' and an auto-mode classifier. Together, these layers make prompt injection virtually impossible to demonstrate.

Rebuilding Claude Code: The 'Press Delete' Philosophy

Cherny revealed that Anthropic adopts a rigorous approach to updating Claude Code. With each new model release, they essentially 'delete all of the codebase, delete all of the prompt and start from scratch every time.' This practice, termed 'ablation' in research, involves deleting the entire system prompt and then reintroducing it line by line to meticulously determine the impact of each component.

This iterative process allows them to adapt the harness to the unique characteristics of each new model. For instance, the intelligence of Opus 5 meant that 80% of the previous system prompt could be deleted, as the model now inherently understands many behaviors that previously required explicit instructions. Cherny encouraged users to experiment with this by setting their own system prompts or using 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 these explicit prompts, though some are still necessary for a user-friendly product experience.

Cherny stressed that this 'press delete' philosophy is crucial for anyone building AI products, emphasizing the need for bravery and a willingness to discard existing structures to accommodate new model capabilities. He also advised users of Claude Code to regularly delete their own configurations (`claude.ai`, skills, hooks) to see how the model behaves with fresh parameters.

Unlocking AI Potential: 'Unhobbling' and Product Overhang

A key concept Cherny introduced is 'unhobbling Claude,' which refers to removing product limitations that prevent models from reaching their full potential. He explained the idea of 'product overhang,' where current models possess unrealized capabilities that existing products do not yet elicit. This is often because the product's design 'hobbles' the model, preventing it from expressing its abilities.

He used the original Claude Code, built on Sonnet 3.5, as an example. At the time, it was an incredible coding model, but existing products offered only single-line autocompletes or read-only chat interactions. The vision for Claude Code was to create a harness that removed these limitations, allowing the model to write entire files and features, thereby unlocking its product overhang.

Cherny believes there's immense opportunity for startups to identify and capture this product overhang with current models. He advised founders to challenge models with tasks that seem slightly beyond their perceived capabilities, describe guardrails and exit criteria, and then let the model 'cook.' He also shared an anecdote about a user discovering Opus 5's ability to draw images using OpenCV, enabling it to create portraits and landscapes simply by being prompted correctly, a capability that was not explicitly trained but emerged organically.

The Future of AI and Engineering

Cherny suggested that coding is becoming 'solved' for many tasks, though challenges remain in highly complex systems code and distributed systems. He posed the question of whether 100% of code can be written by agents, noting that while the desktop app is now quite fast, the journey is ongoing.

He concluded by emphasizing that the most effective users of Claude are those who approach it empirically, forgetting prior knowledge of older models and focusing on understanding the current model's behavior through experimentation. This scientific mindset, he argued, is key to unlocking the full potential of AI.

For those studying computer science before the era of AI agent coding, Cherny advised learning practical application, whether through startups or personal projects, rather than solely focusing on theoretical knowledge. He suggested that combining computer science skills with business acumen, user understanding, and data science knowledge is where true value lies.

Finally, Cherny announced a special offer for attendees: max 20x compute credits for Claude, encouraging them to experiment and build the next generation of AI applications.