Kieran Klaassen's 'Compound Engineering': Building Products with AI and Zero Code

This product manages email inbox functionality, a complex domain, and represents the culmination of two years of dedicated effort. Klaassen's method challenges traditional notions of product development by minimizing direct coding.

Cora is designed to be highly versatile, supporting various interfaces including a command-line interface (CLI) and a desktop application. This flexibility underscores the power of agent-native design in AI-driven products.

As a founder and former VP of Engineering, Klaassen intentionally structured the Cora project to push the boundaries of AI-assisted development by a single engineer. He leveraged AI models like Sonnet 3.5 with the explicit goal of evaluating how far AI could extend his capacity before needing to expand the team. This experimental approach aimed to determine the maximum individual output using advanced AI tools.

Klaassen monitored the shifting bottlenecks in his development process over two years as AI capabilities improved. Initially, code generation was the primary hurdle, but this quickly shifted to planning as AI became more proficient. Eventually, the core challenge evolved to determining 'what to build' rather than 'how to build it.' This progression led to the essential development of a memory system to prevent the repetition of instructions and ensure efficient AI learning.

Klaassen observes a fundamental shift in modern engineering, moving away from direct typing and implementation towards higher-level judgment and taste. He believes that AI has largely solved the implementation phase, making the human role more about providing strategic direction and critical decision-making. The compounding system he developed allows AI to retain previous brainstorming results, reducing repetitive tasks and enabling engineers to focus on qualitative aspects.

This loop involves brainstorming, planning, working, reviewing, polishing, compounding, and repeating. He asserts that single engineers leveraging these compounding systems can achieve more than larger teams without such a framework.

He suggests that 50% of the time should be dedicated to building the feature and ensuring it delivers the intended value. The other 50% should be spent on teaching the system, specifically correcting any mistakes it made. This teaching process is crucial for improving the AI's future performance and overall system efficiency.

Klaassen explains that he stores all acquired knowledge as solution documents within his repository. He argues that this approach is more token-efficient than relying on AI for deep research, despite common perceptions about token usage. By having the correct answers and solutions readily available, the system reduces the need for extensive review, correction, or redundant research, ultimately saving costs and improving efficiency.

Klaassen has developed a 'Compound Engineering plugin' that users can install in various tools, such as CodeExCloud and Cursor. This plugin extends the benefits of his workflow to a broader audience, including hundreds of thousands of daily users. It is designed to assist engineers, product managers, and designers in integrating compound engineering principles into their daily operations.

The 'C ideate' command is designed to transform chaotic backlog data from various sources like Linear, GitHub, Slack, and Intercom into structured arguments. This tool analyzes the feedback and generates reasoned arguments about which tasks are worthwhile to pursue versus those that are not. The output is a clean HTML document, suitable for sharing with the team, thereby streamlining backlog management and strategic planning.

The system allows users to align ideation with their strategic objectives by pointing it to existing Objectives and Key Results (OKRs). It utilizes past learnings and strategy documents to ensure that new ideas are consistent with the overall direction. The results can then be exported to generate presentations, making it easier to communicate and justify strategic decisions.

The system provides sharp, critical questions, which he finds particularly valuable. After answering these questions, the knowledge is 'compounded' back into the system, enabling it to learn and avoid asking the same questions in future reviews. This iterative process refines the system's understanding and improves documentation quality.

This command acts as a 'brain on' prompt, signaling the need for dedicated, uninterrupted work on intricate problems.

Klaassen emphasizes that his intelligent query system balances inquiry with autonomy, adapting its communication based on user personas and specific task requirements. The goal is to achieve the best work outcome, not to generate an excessive number of questions. He notes that many AI libraries fall into the pitfall of asking too many questions, which can hinder efficiency rather than help.

This system can run for hours, performing tasks like planning, work review, and testing. It automatically opens a pull request, conducts dogfooding, attempts to fix issues, and generates 'before and after' video screenshots within the pull request, even operating overnight and in parallel to other tasks.

To maintain high standards, the user reviews the product alongside the code. When specific UI errors, such as duplicate logos, are identified through human feedback, that knowledge is extracted and fed back into the system. This process ensures that the AI learns from past mistakes. Consequently, future design work is correctly tagged, the relevant files are found, and the system knows to avoid repeating those errors, effectively 'closing the loop' on the learning process.

Klaassen advises developers to never repeat the same manual task, urging them to extract the underlying logic into the system to prevent recurrence. He emphasizes the importance of creating an autonomous middle layer for routine operations. The strategy involves manually iterating on a process until it becomes reliable enough to run unsupervised, thereby compounding knowledge and eliminating repetitive work.

Klaassen's strategic outlook posits that implementation costs will continue to decline, while expert judgment will remain invaluable. He believes that future models and systems must be designed to access and leverage human judgment effectively. Engineers, therefore, must own the problem definition at the start of the process and continuously raise the bar for quality at the end. Success, in this context, is defined by ensuring that each subsequent feature becomes easier to build than the last.