Boris Cherny: Building Claude Code

Opus 5 can run for days, weeks, or months without stopping, even without external scaffolding like LangGraph (SLGO). When combined with auto mode, it sustains long tasks autonomously, a capability no other model has demonstrated. This is a key advancement for agentic workflows.

Boris Cherny states Opus 5 is no longer prompt injectable, a breakthrough built on three layers: a well-aligned model (three years of alignment research), a prompt injection classifier based on Chris Olah's mechanistic interpretability that monitors neurons lighting up during injection attempts, and an auto mode classifier. Together, they make prompt injection undemonstrable.

With each new model, the Claude Code team deletes and rebuilds the system prompt from scratch. For Opus 5, they removed 80% of the prompt because the model is intelligent enough to handle tasks without many prior corrections. A 'simple mode' (CLAUDE_CODE_SIMPLE=1) deletes all prompts including tool prompts, and surprisingly the model is slightly more intelligent without them, though some prompts remain for product usability.

Boris Cherny advises users to avoid overly specific instructions when using Claude Code. Instead, he recommends describing the task, guardrails, and exit criteria at a high level, then letting the model work autonomously. This approach, which would not have worked six months ago, now yields surprising results due to model improvements.

The Bun team used Claude Code to rewrite their entire JavaScript runtime from Zig to Rust in a single dynamic workflow that ran for 11 days. The model was given one prompt with steering, leveraging Bun's extensive test suite to validate correctness. This task, which would have taken over a year for the best engineers, is now in production and powers Claude Code itself.

Cherny introduces the concept of 'product overhang'—continuously re-testing the latest model on previously unsolved business or engineering problems, as new models may succeed where older ones failed. He emphasizes that users should experiment freely, as capabilities emerge without explicit training, such as Opus 5's ability to draw using OpenCV, discovered accidentally.

Cherny notes that the skill of eliciting model capabilities is evolving from 'prompt engineering' to 'context engineering.' He suggests that the focus is now on providing the right context and framing rather than crafting intricate prompts, as models become more capable of understanding high-level instructions.

Boris Cherny clarifies that coding is solved for the kind of coding he does, but not for everyone. He notes that Claude still struggles with super deep systems code bases, distributed systems, and in-the-weeds UI verification like pixel-perfect checks. He highlights that Opus 5 was a big leap in vision and computer use but is still not perfect, and he polls the audience on how much of their code is written using agents, finding that many use it for over 50% of their work.

Cherny explains that the best users of Claude bring an empirical mindset: they forget past model behaviors and computer science theory, instead looking at the model, trying a task, seeing where it struggles, and adjusting based on that. He emphasizes that this is no longer a theoretical science but an empirical one, and the key skill is being able to let go of priors and remain open to trying again, which is now very successful.

When asked what students should still learn the hard way, Cherny shares his personal story of learning to code on a TI-83 calculator in middle school to cheat on math tests, starting with BASIC and later learning assembly for calculus. He advises students to learn not just computer science theory but also how to apply it by building startups, products, developing design and business sense, doing data science, and talking to users, as combining these with engineering is where true value lies.

The interviewer summarizes Cherny's advice as starting with making something for yourself first, then leveling up to make something people want. Cherny agrees, and then announces a special offer for the audience: they are getting max 20x credits, encouraging them to build something that runs for multiple months with thousands of agents.