Compound Engineering: Building Cora Without Code

Kieran Klaassen, the solo developer behind the email client Cora, has a radical approach to software development. This year, he hasn't written a single line of code, yet Cora serves thousands of users. His journey, detailed in a recent discussion, reveals a shift in how AI can be integrated into the development process, moving beyond traditional coding to a more collaborative, teaching-centric model.

The Evolution of Cora's Development

Klaassen's work on Cora began with a focus on code quality. Two years ago, the codebase was a significant constraint. He addressed this by layering on review processes and improving his own skills until the code reached a satisfactory standard. The next bottleneck became the planning phase. Once the plans were refined, the challenge shifted to knowing what to build. This led him to develop a system for managing and refining product direction.

The final hurdle was the inherent repetition in his own work. This is where the concept of a memory system, and subsequently compound engineering, emerged. The core idea is to fix the human tendency to repeat mistakes or inefficiencies.

Compound Engineering: A New Development Paradigm

Compound engineering, as defined by Klaassen, is a demanding yet efficient methodology. It requires developers to spend half their time building a feature and the other half teaching the AI system what it did incorrectly during the feature's development. This might sound counterintuitive, potentially increasing costs. However, Klaassen argues it's cheaper in the long run, especially in terms of 'tokens', a reference to the computational resources used by AI models.

"His counterintuitive claim is that this ends up cheaper in tokens rather than more expensive, because a stored solution means no correction pass and no research detour the next time around."

This loop places the human at both ends of the development process. The initial phase requires the human brain to clearly define the problem and the desired outcome. Then, after hours of autonomous AI work, the human brain re-engages not for quality assurance testing, but to elevate the system's performance. The ultimate goal is that the next feature should be easier to build because the previous one was successfully shipped and learned from.

From Backlog to Argued Ideas

Klaassen also discussed how he transformed a traditional backlog into a more dynamic and argumentative set of ideas. This involves asking sharp, precise questions about documents and ideas, ensuring only the most critical aspects are considered. This iterative questioning process helps refine concepts before they are handed off for AI development.

The overnight loop is a key component of this process. It allows for extensive autonomous work by the AI, with the human developer reviewing and refining the output each morning. This polish, Klaassen notes, is not just about fixing errors but about actively raising the bar for future development. The objective is a system where complexity does not accumulate in a negative way, but rather, each completed feature makes the next one simpler to build.

StartupHub.ai data shows that companies focused on developer tools and AI productivity, like Cora is building towards, often face initial low scores due to their nascent stage. For instance, January, a tool that helps manage AI development, has a StartupHub score of 2/100. Two, another AI-focused company, scores 25/100, indicating the early-stage challenges and opportunities in this emerging sector.

The Cora Plugin and Future Implications

The development of Cora's plugin was integrated directly into the product-building process. This meant building the necessary tools and infrastructure as the core product evolved. This approach ensures that the tools are directly relevant and optimized for the tasks at hand.

Klaassen's ability to ship a functional email client without writing code highlights a potential future for software development. It suggests a move towards more human-AI collaboration, where human ingenuity is focused on problem definition and learning, while AI handles the execution and iteration. The principle of compound engineering, where each shipped feature makes subsequent development easier, could fundamentally alter how software complexity is managed and how efficiently products are built.