Boris Cherny Unveils Claude Code: A New Paradigm for AI-Driven Software Development

Early coding models primarily focused on single-line autocomplete functionalities, offering limited assistance to developers. Boris Cherny noted that these models often lacked comprehensive read-write access to entire codebases, restricting their utility. Claude Code was developed with the explicit goal of eliminating such scaffolding, thereby enabling models to edit and generate entire files or functions at once, unlocking a new level of autonomous coding.

Cherny stated, "the feeling was that there wasn't really a product that was fully eliciting the model's capability to write entire functions at a time, entire files at a time."

Modern AI models possess a vast array of untapped capabilities that existing products frequently fail to fully utilize, a phenomenon Boris Cherny refers to as "product overhang." He suggested that startups have a significant opportunity to create value by designing interfaces and platforms that 'unhobble' these models, allowing them to perform at their full potential. Claude Code's success, in this regard, stems from its provision of full terminal access, which enables models to interact with the environment in a more comprehensive and unconstrained manner.

Cherny noted, "I think that nowadays with modern models there is so much product overhang that I I'm not saying startups capture and I think there's people thinking about these problems but there's just a huge amount amount of opportunity to elicit these behaviors from the model that are just like amazing and interesting and and commercially valuable."

When prompting modern AI models, Boris Cherny advises moving away from overly specific, step-by-step instructions. Instead, users should adopt a higher-level approach by describing the overall task, defining guardrails, and clearly stating the exit criteria for successful completion. This method allows the model more autonomy to "cook" and often leads to surprising and effective results, leveraging its advanced problem-solving capabilities.

Cherny explained, "You want to go a little bit higher level. You want to describe the task, you want to describe the guardrails, you want to describe like the exit criteria and then just go let the model cook and come back in a little bit and I think it'll it'll surprise you"

Boris Cherny presented a compelling case study involving the migration of Bun, a JavaScript runtime, from Zig to Rust using Claude Code. This extensive task was initiated with a single prompt and executed through dynamic agent workflows, a feature in Claude Code designed to orchestrate numerous agents for productive work. The automated process ran for 11 days, successfully rewriting the entire codebase.

This undertaking highlighted the profound impact AI agents can have on large-scale engineering projects, potentially reducing years of human effort to mere days. The use of dynamic workflows allowed for the sophisticated orchestration of dozens, hundreds, or even thousands of agents to complete the migration efficiently.

Cherny described the feat, "he had the model rewrite it from Zig to Rust. It was one prompt. It was a dynamic workflow. And a dynamic referrals are a feature in cloud code that essentially let you orchestrate, you know, dozens, hundred, thousands of agents to do work productively. And it ran for 11 days and it rewrote the entire codebase."

Claude Code is currently running in production environments, demonstrating its capability to manage highly complex software tasks. It is utilized for projects involving over 100,000 lines of JavaScript runtime code, validating its robustness and scalability in real-world scenarios.

Boris Chernyshev confirmed its active deployment, stating, "This is in production out. This is what Claude Code uses now when you're running it."

Boris Chernyshev emphasized the importance of continuously re-testing difficult engineering or business problems with the latest available models, even if previous versions failed. He pointed out that advancements in new models can enable them to succeed where older ones could not. This iterative approach ensures that users leverage the most current AI capabilities for challenging tasks.

Chernyshev advised, "You should just keep throwing the latest model at it to see if it'll just do it because even if a previous model didn't, the new one might."

Boris Chernyshev highlighted a significant shift in the core skill required for effective AI interaction: from traditional prompt engineering to robust verification. The key is now to design tasks where the model can not only attempt a challenging problem but also verify its own work throughout the process. He identified verification as the most overlooked, yet crucial, element for achieving successful outcomes with AI.

Chernyshev explained, "I think the skill nowadays is less about prompt engineering and more about figuring out how do you give Claude a hard task that seems a little bit too hard and then how do you make it possible for Claude to verify its work along the way"

In another case study, Boris Chernyshev tasked Claude with a complete rewrite of an Electron application into Swift. This ambitious project requires a Mac virtual machine for the AI agent to run and test the Electron app. Claude then compares screenshots of the Electron and Swift versions pixel by pixel, iteratively refining its work until the visual output is identical, ensuring a high degree of accuracy and completion.

Chernyshev outlined the instructions, "I want you to rewrite the Electron app in Swift. I want you to run the Electron app in the Mac virtual machine, screenshot it, and then look pixel by pixel, compare it to the Swift version, don't stop until you're done."

The Swift rewrite task, as described by Boris Chernyshev, has been in progress for over two weeks, running autonomously for approximately 14 to 15 days. During this extended period, Claude independently reports its progress through screenshots sent via internal Slack channels, demonstrating its capability to manage long-term projects without constant human oversight.

Chernyshev noted the duration, "It's been uh it's been a little over two weeks. So, it's like 14 days, 15 days."

Boris Chernyshev dismissed the notion of a single "one weird trick" for effective AI interaction, emphasizing instead an empirical approach. He argued that success comes from iteratively engaging with the model: assigning it a task that initially seems too difficult, providing it with tools for self-verification, and then refining the process when the model encounters difficulties. This method focuses on hands-on problem-solving and continuous improvement.

Chernyshev stated, "I think everyone's looking for like the one weird trick to do it. The there's just like that doesn't exist. The way the model works is you have to approach it empirically. You have to give it a task that's too hard. You have to give it the tools to verify the work."

Boris Chernitsky observed that experienced engineers often overengineer AI tasks by attempting to specify them too precisely, treating AI models like traditional code. This approach, rooted in past system-building methodologies, can hinder the model's ability to operate effectively. He suggested that the model performs best when treated as a collaborative coworker rather than a rigid script, implying a need for engineers to unlearn some of their established habits.

Chernitsky commented, "I think people tend to overengineer because I think in a lot of ways like when we built systems in the past, that's the way you had to do it."

Boris Chernitsky estimates that Claude Code can spawn thousands to tens of thousands of agents to handle complex, large-scale engineering tasks. He noted that the most effective users leverage these high agent counts to gain significant leverage, enabling them to tackle work that would be impractical with fewer resources. This massive orchestration capability is central to Claude Code's ability to automate extensive projects.

Boris Chernitsky introduced dynamic workflows as a new form of "test-time compute," distinguishing it from traditional scaling laws based on training data or model size. Test-time compute refers to the volume of tokens generated during the inference phase, and dynamic workflows provide a novel method to scale this compute effectively. This allows models to solve increasingly difficult tasks by orchestrating computational efforts during their execution.

Chernitsky articulated, "This is actually like a new form of test time compute. ... And now dynamic workflows are essentially a new way to orchestrate test time compute."

Boris Chernitsky outlined two key features for persistent, scheduled tasks: 'loops' and 'routines.' Loops are comparable to a cron job running locally for Claude, while routines serve the same purpose but execute in the cloud, ensuring persistence even if the user's laptop is closed. These features are valuable for tasks that don't share context but may accumulate shared memory over time, allowing them to be scheduled hourly, every five minutes, or daily for continuous operation.

Chernitsky clarified, "Loop is essentially a cron job that's running locally for quad. Routine is the same thing, but it's running in the quad in the cloud."

Claude Code employs automated routines to maintain its own codebase, a process Boris Chernitsky described as 'Claude maintaining itself.' This includes daily execution of tasks such as cleaning up dead code, running tests, and shipping experiments, all initiated by single-sentence prompts. The system also features an 'abstraction police' routine designed to identify and consolidate duplicate code patterns across various platforms, including CLI, iOS, Android, and desktop applications.

These self-maintenance capabilities ensure code consistency and efficiency, significantly reducing the manual effort typically required for continuous integration and development. Chernitsky elaborated, "We actually have quad maintaining itself now. ... One routine is clean up dead code. This is a single prompt. It's like one sentence. Cloud runs this every day."

Despite significant advancements, AI agents, including Claude, still face limitations in complex software development, particularly with deep systems code and intricate distributed systems. Boris acknowledged that the problem of coding is "not solved for everyone," citing struggles with highly specialized codebases. Additionally, UI verification remains a challenge, as achieving pixel-perfect accuracy in user interfaces is difficult even with improvements in vision capabilities, such as those seen in Opus 5.

Boris advocated for an educational approach that combines theoretical computer science with practical application. He shared his personal journey of learning to code by solving real-world problems, such as automating math test solutions, which drove his evolution from BASIC to assembly language. He advises students to not only understand the intellectual fascination of computer science but also to learn how to apply it effectively.

He recommends that students integrate business acumen, design principles, data science, and effective user communication into their skill set. This holistic approach prepares individuals not just for coding, but for creating valuable solutions in diverse contexts.

Boris advised, "learn not just the computer science. This is like intellectually fascinating and it's really really interesting to know but learn how to apply it."

To conclude the event, Boris announced a special gift for the audience: 20x access to Claude Code for their personal projects. This generous offer aims to enable participants to directly experiment with the platform's advanced AI coding capabilities.