Best of the Pod: Claude Code - How Two Engineers Ship Like a Team of 15
Timestamps are as accurate as they can be but may be slightly off. We encourage you to listen to the full context.
Podcast Summary
In this dynamic episode of AI & I, host Dan Shipper interviews Kieran Klaassen (GM of Cora) and Nityesh Agarwal (Cora engineer) about their revolutionary approach to engineering with AI agents. (01:29) The duo shares how they've transformed their development process using Claude Code, creating what feels like a 15-person team with just two engineers. (02:58) They reveal their breakthrough realization that AI coding should extend far beyond just writing code to encompass research, workflow management, and comprehensive task automation.
- Main themes: The evolution from traditional AI-assisted coding to full agentic development workflows, the compound effect of building prompts that create other prompts, and practical strategies for managing AI agents in production environments.
Speakers
Kieran Klaassen
General Manager of Cora, Every's AI email assistant, and resident AI agent aficionado. Kieran is an experienced Rails developer who has become a leading voice in agentic coding workflows. He's known for his innovative approach to voice-to-text programming and his comprehensive testing of AI coding tools, having ranked and used virtually every available AI agent in the market.
Nityesh Agarwal
Software Engineer at Cora who represents the new generation of AI-native developers. Nityesh taught himself programming using ChatGPT two years ago and has evolved through every major AI coding tool transition. He brings fresh perspectives on AI-first development workflows and draws inspiration from management principles like those found in "High Output Management" to optimize AI agent processes.
Key Takeaways
Expand AI Usage Beyond Coding to Full Workflow Management
The biggest breakthrough comes from realizing that coding with AI should encompass research, planning, project management, and everything surrounding development—not just writing code. (04:24) Kieran emphasizes that agents are now capable enough to handle comprehensive workflows, requiring a complete rethink of how developers approach their work. This means using AI for tasks like analyzing Git logs, understanding user feedback, creating detailed project specifications, and managing GitHub issues. The compound effect occurs when each AI-powered task makes subsequent tasks easier and faster to execute.
Build Prompts That Generate Other Prompts for Compounding Effects
The most powerful strategy involves creating meta-prompts that generate other prompts, creating a compounding engineering effect. (23:07) Kieran and Nityesh developed a custom command in Claude Code that takes a simple feature idea and automatically generates comprehensive research documents, including problem statements, technical requirements, implementation steps, and best practices research. This approach dramatically reduces the manual work needed for each new feature while maintaining consistency and thoroughness in planning.
Implement Human Review at the Lowest Value Stage
Drawing from Andy Grove's "High Output Management," the key principle is to catch and fix problems at the earliest, lowest-cost stage of the development process. (33:44) This means investing time in reviewing AI-generated plans and specifications before allowing agents to implement solutions. Nityesh emphasizes that while it's tempting to let AI immediately start coding after generating a plan, catching directional errors early prevents expensive rework later in the development cycle.
Use Traditional Testing and Evaluation Methods as Safety Nets
Despite the advanced capabilities of AI agents, traditional software engineering practices like testing and evaluation remain crucial for maintaining quality. (39:12) Kieran recommends implementing smoke tests, automated evals for prompts, and systematic testing protocols. He describes using Claude Code to run evaluations multiple times, identify failure patterns, and iteratively improve prompts until they pass consistently—all while he grabbed a coffee, demonstrating the autonomous nature of modern AI workflows.
Leverage Multiple Specialized Agents for Different Tasks
Rather than relying on a single AI tool, the most effective approach involves using different agents for their specialized strengths. (47:58) Kieran uses Friday for UI work, Claude Code for research and complex workflows, and Charlie for code reviews. This ecosystem approach mirrors hiring specialists for different roles, allowing each agent to operate in its area of maximum effectiveness while maintaining standard development workflows that human team members can easily integrate with.
Statistics & Facts
- The Cora team shipped six major features, five important bug fixes, and three infrastructure updates in a single week with just two engineers. (09:57) This demonstrates the dramatic productivity gains possible with agentic coding workflows.
- Kieran successfully ran Claude Code for 25 minutes continuously on a single complex task, while Nityesh achieved 8 minutes, showing the autonomous capabilities of modern AI agents. (22:23)
- The team created approximately 20 detailed GitHub issues in a single brainstorming session using their automated workflow, demonstrating the speed of AI-powered project planning. (27:02)
