Claude Code as the Engineer

Claude Code as the Engineer is my operational model for software development, where I delegate core implementation tasks to an AI agent and retain the role of a senior reviewer and architect.

What it is

This approach treats Claude Code not as a co-pilot, but as a dedicated implementation engineer. My role shifts from writing the majority of the code to defining the architecture, setting the technical specifications, and then performing a rigorous staff-level code review. Claude Code handles the boilerplate, the CRUD operations, the API integrations, and the component scaffolding across my various product lines at Total Ventures. It's about leveraging an agent to handle the high-volume, predictable coding tasks, freeing me to focus on strategic decisions and system integrity. I provide the requirements, often in natural language or with minimal code examples, and Claude Code delivers a pull request.

Why I do it this way

The core driver is leverage. Running a multi-product studio with AI as the team means I cannot be the bottleneck for every line of code. This model allows me to manage the development of projects like Total Formula 1, Pregnancy Power Hour, and Inky concurrently, without sacrificing quality or consistency. It's an application of Multiple Revenue Streams, One Back Office at the code level — a single "engineer" (Claude Code) can contribute across diverse projects, maintaining a consistent style and structure I've defined. I learned the hard way that trying to author every piece of every system myself is unsustainable. This approach enables me to ship more, faster, by focusing my human attention where it adds the most value: design, review, and complex problem-solving.

How it works in practice

My workflow begins with a detailed specification for a feature or component. For example, building a new payment flow for Inky might involve defining the Stripe integration points, the database schema in Firebase, and the email notifications via Resend. I feed these requirements to Claude Code, often with existing code snippets or architectural diagrams from my monorepo. Claude Code then generates the necessary files, functions, and tests, typically submitting them as a pull request to the relevant repository. I review every PR. This isn't a rubber stamp; it's a critical review, much like a staff engineer would scrutinize a junior's work. I fix the 5% Claude got wrong or where it made an inefficient choice, and then ship the 95% it got right. This often involves minor refactoring, optimizing database queries, or adjusting an edge case in a UI component for Pregnancy Power Hour. For new features like the Lead Magnet → Paid Ladder implementation for Total Formula 1, Claude Code handles the initial data capture forms and API calls, while I ensure the conversion logic is sound. Consistency is also a key benefit; by loading a Voice as Source of Truth for each brand, Claude Code can even generate content and code comments that align with the brand's specific tone.

Where this breaks down

While highly effective for established patterns and well-defined tasks, Claude Code struggles with truly novel architectural problems or deeply complex, ambiguous requirements. When a problem requires abstract reasoning, creative problem-solving outside of known patterns, or debugging intricate race conditions, human intervention is essential. It also performs poorly when the initial prompt is vague or contradictory; garbage in, garbage out applies rigorously here. I've found that debugging Claude Code's errors can sometimes take longer than writing simple functions myself, especially if the error stems from a fundamental misunderstanding of the system's intent. This is where my role as the staff engineer becomes critical—identifying when to let the agent build and when to step in and architect or debug manually.