Three Rules for Building Projects with Claude Code

Avoid the Idea Trap

Most projects fail because developers either lack a clear user base or attempt to build tools that compete directly with frontier AI labs. If you are the sole user, prioritize speed and function over aesthetics. If you are building for others, you must identify a specific, narrow set of users and solve their problems elegantly. Avoid building general-purpose AI tools or security auditing software, as these are areas where major labs are already deploying massive resources. Before starting, ask: can I name five specific people who would use this today, and does this project become more valuable as AI models improve?

Build Where You Live

Your competitive advantage is not your ability to prompt AI, but your domain-specific judgment. Use the T-shaped model: your surface-level knowledge is the horizontal bar, but your deep, practical experience—knowing what works and what fails in a specific industry—is the vertical bar. Focus your efforts on the vertical bar where you have earned expertise, such as healthcare, legal, or education, rather than competing in saturated general-purpose categories.

Operate as a CEO

Shift from the execution layer to the leadership layer by orchestrating AI rather than doing the work yourself. Treat Claude Code as a new hire by implementing these six operational moves:

• Create a claude.md file to serve as an onboarding document, providing the AI with necessary context to reduce correction cycles.
• Require the AI to interview you before writing code to define the core problem, success metrics, and non-goals.
• Configure permissioning to allow agents to perform reversible actions autonomously while requiring manual approval for destructive tasks.
• Build a cabinet of specialized experts by training agents on specific playbooks, such as sales, content, or finance.
• Review outputs like a manager: have the AI generate multiple options and select the best one rather than asking for end-to-end solutions.
• Remove yourself as a bottleneck by utilizing power-user features like hooks for session logging, scheduled agents for recurring tasks, and loops for automated system maintenance.