Stop Hobbling Your AI: unleash latent capabilities with Higher-Level tasks

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 The central thesis is that current AI models possess significant unrealized capabilities due to restrictive product design and overly specific instructions. The clusters detail how Opus 5's advancements in prompt injection resistance enable safer agent design.

Opus 5 has largely overcome prompt injection vulnerabilities, a critical security flaw that previously allowed malicious external instructions to override a model's intended behavior. This enhanced safety means developers can build more reliable and secure AI agents.

Concurrently, models like Opus 5 demonstrate remarkable capabilities, such as rewriting entire codebases between languages in days, a task that previously required years of human effort.

For founders and developers, improved prompt injection resistance drastically reduces security risks, enabling the creation of more trustworthy and scalable AI products. The ability to automate complex engineering tasks like code migration offers immense value, accelerating product development and modernization efforts, and opening doors to previously unfeasible projects.

Developers should prioritize leveraging Opus 5's enhanced security for agent design and explore its capabilities for large-scale code transformation projects. This allows for faster iteration, reduced technical debt, and the ability to tackle ambitious engineering challenges with unprecedented efficiency.

Models often possess capabilities that are not immediately obvious or were not explicitly part of their training data. Through creative experimentation and a willingness to 'play' with the model, users can discover these hidden talents.

This 'solicitation gap' highlights that the model's ability to perform a task depends heavily on how it is asked, not just what it was trained on.

For innovators, this means that the potential of AI models extends beyond their documented features. Fostering a culture of playful exploration can lead to groundbreaking discoveries and novel applications, giving early adopters a significant competitive advantage by unlocking unique functionalities.

Organizations should allocate resources and time for 'AI play days' or hackathons, encouraging engineers and product teams to experiment freely with models. This can lead to unexpected product features, efficiency gains, or entirely new business opportunities.

AI's capabilities in coding are advancing rapidly, solving many application-level tasks. However, complex areas like deep systems programming, distributed systems architecture, and precise UI/UX verification still present significant challenges for AI.

This means human developers will shift their focus to these more intricate problems. The most valuable skills for future computer scientists will be the ability to apply technical knowledge to real-world problems, combined with a strong understanding of business, design, and user needs, moving beyond purely theoretical computer science.

This insight is vital for founders, leaders, and educators. It clarifies where human talent remains indispensable and guides the development of future skills.

Investing in interdisciplinary education and fostering problem-solving mindsets will be key to building successful teams and products in an AI-augmented world.

Founders should strategically deploy AI for routine coding tasks, freeing human engineers to focus on complex systems, architectural challenges, and nuanced user experience. Educational institutions should adapt curricula to emphasize practical application, interdisciplinary skills, and problem-solving over rote technical knowledge.