The Era of Compound Engineering: Building Cora Without Writing a Single Line of Code

Kieran Klaassen, the mind behind the email client Cora, states that he has not written a single line of code this year, nor has he extensively reviewed it.

Despite this, he successfully ships his product, which thousands of users rely on for their email inboxes, attributing this capability to a sophisticated system designed to capture his expertise and preferences.

Kieran Klaassen defines compound engineering as a method he developed while creating Cora, an AI-native email inbox.

This approach is characterized by leveraging accumulated knowledge and operating in fast feedback loops, with Cora functioning as an agent-native system that runs across desktop, phone, and command-line interfaces, effectively shifting the focus from manual typing to applying judgment and taste.

Klaassen rebuilt the entire Cora email client by himself, starting in January, using Ruby on Rails for the backend and React for the frontend.

He intentionally avoided hiring additional team members to rigorously test the operational limits and capabilities of advanced AI models such as Sonnet 3.5.

Two years ago, when Klaassen began, the primary bottleneck was poor code quality and AI hallucinations, which he resolved by implementing agent-based systems and rigorous review processes.

The subsequent challenge was planning, addressed by leveraging AI to generate comprehensive project plans, and then finding product-market fit, which he tackled by becoming a dedicated user of his own product.

The final bottleneck identified was memory, which led to the creation of a system designed to store and leverage accumulated knowledge, thereby preventing repetitive issues.

He employs a "human-AI sandwich" model, where the human initiates and concludes the development loop, providing the essential contextual framework and final approval.

This approach ensures that the human intellect is focused on the critical areas where models cannot replicate nuanced understanding and strategic direction.

The other half should be invested in teaching and refining the system based on any errors or shortcomings it encountered, with the objective of significantly improving the system's performance in subsequent iterations.

Klaassen explains his practice of storing accumulated knowledge within his repository as structured solution documents, which, contrary to common concerns about token usage, actually proves to be more token-efficient in the long run.

By having pre-baked solutions and accurate answers directly in the context, the system drastically reduces the need for repetitive deep research, extensive reviews, and continuous corrections, streamlining the development process.

Developed with the assistance of collaborator Trevan Chowo, who helped refine its product sense, the plugin is designed for a broad audience including engineers, product managers, designers, and general knowledge workers seeking to compound their expertise.

The CE ID8 tool automatically processes and structures issues from various sources, including Linear, GitHub, Slack, and Intercom, transforming fragmented information into organized insights.

It generates clear HTML strategy pages and provides arguments for which tasks are most beneficial to work on, aligning them with predefined Objectives and Key Results (OKRs).

This allows the system to score ideas based on their strategic importance, ensuring that development efforts are consistently directed toward high-value work.

Klaassen recommends using the doc review feature on Product Requirements Documents (PRDs) or other project documents to elicit sharp, insightful questions.

The answers to these questions are subsequently compounded back into the system, which enables the AI to learn from the context and avoid asking the same questions repeatedly in the future.

The CE brainstorm tool is specifically designed for handling large-scale tasks, such as version upgrades, by pulling in compounded knowledge from the repository.

It integrates defined personas to refine the output, ensuring that the brainstorming process is grounded in comprehensive context and diverse perspectives.

Klaassen emphasizes that the goal of AI-assisted workflows is to achieve the absolute best work, rather than merely generating a high volume of interactions or questions.

He advises against excessive questioning and instead advocates for dialing in the system to ask precisely the right amount of questions, ensuring that every interaction directly contributes to a superior outcome.

Klaassen describes LFG as an extensive automation loop capable of running for hours, handling planning, work review, and testing, culminating in the opening of a Pull Request (PR).

The system also performs dogfooding, automatically identifies and fixes issues, and generates before-and-after video screenshots within the PR, enabling parallel processing and asynchronous work.

Klaassen stresses that the aim is not merely quality assurance, but to fundamentally raise the bar for what is considered acceptable output, asserting that if the system fails, the LFG flow itself has failed.

When an error occurs, the user provides corrective feedback, which the system then incorporates into its knowledge base, thereby preventing the recurrence of similar mistakes in future iterations.

Compound engineering fundamentally involves constructing autonomous systems that continuously run, plan, and review tasks, making development processes significantly less mundane.

The core principle is to identify and extract any repetitive work into the system, ensuring that such tasks are never repeated manually, thereby continuously improving efficiency.

Moreover, documentation in this paradigm should prioritize the underlying reasoning behind the code, rather than solely focusing on the code itself, to facilitate deeper learning and prevent future errors.

Klaassen argues that while implementation costs are decreasing, human judgment and taste remain the primary bottlenecks in software development.

The goal of compound engineering is to fundamentally alter the paradigm so that each new feature becomes easier to build than the last, rather than increasing complexity as is traditionally the case.

This shift entails leveraging AI to handle the heavy lifting of development, allowing humans to focus entirely on applying their judgment and strategic direction.