The Philosophy
I don't write code from scratch anymore. I architect systems, write prompts, and let AI agents do the heavy lifting. Over the past 2 years, I've built 90+ projects using this approach — and it's not about being lazy. It's about leverage.
The Core Loop
My development process follows a tight loop:
- Research — I use AI agents to explore the problem space, gather requirements, and identify patterns from similar projects
- Architecture — I design the system structure, data models, and API contracts myself. This is where human judgment matters most
- Scaffolding — AI agents generate the boilerplate, file structures, and initial implementations
- Iteration — I review, refine, and direct agents to fix issues, add features, and optimize
- Ship — CI/CD pipelines, testing, and deployment happen with minimal manual intervention
Skills and Sub-Agents
The real power comes from structuring AI workflows as skills — reusable, composable units of capability — and sub-agents — specialized agents that handle specific domains like:
- Research agent — Gathers information, summarizes documentation, compares alternatives
- Architecture agent — Suggests patterns, reviews designs, identifies edge cases
- Implementation agent — Writes code, generates tests, creates documentation
- Review agent — Audits code quality, security, performance, and accessibility
Prompt Engineering is the New Coding
The skill isn't typing fast. It's asking the right questions, structuring the right constraints, and knowing when to trust the AI versus when to intervene.
Key Takeaways
- Start with architecture, not implementation. AI can write code but can't (yet) make good architectural decisions
- Build a library of reusable prompts for common patterns
- Use agents for repetitive tasks: scaffolding, tests, documentation
- Always review AI-generated code — it's your responsibility
- Ship fast, iterate faster. The goal is learning, not perfection