The Philosophy of the 'Builder' Workflow

Building custom software is no longer about mastering syntax; it is about domain expertise and clear communication. The most common pitfall for builders is jumping straight into code without a clear roadmap, leading to bloated, disconnected tools. The goal is to build lean, purpose-built utilities that solve specific workflow gaps—like bridging the gap between time tracking and invoicing—without the overhead of generic SaaS platforms.

Shaping the Product Requirements Document (PRD)

The process begins with a 'brain dump' of high-level features. Using a PRD creator skill, the developer iterates with the AI to refine scope. Crucially, this phase is used to define what is out of scope to maintain a lean V1. The PRD acts as the 'source of truth' within the codebase, allowing for course corrections as the project evolves.

Key to this stage is the interview process: the AI should be prompted to ask clarifying questions about UI, UX, and data modeling. By forcing these decisions early, you avoid technical debt and logic conflicts (e.g., deciding how to handle time entries that have already been invoiced) before a single line of code is written.

Milestone-Based Execution

Once the PRD is locked, the project is broken into logical, buildable milestones. This prevents the AI from becoming overwhelmed or losing context. The workflow follows a strict pattern:

  1. Plan Mode: Before coding, the AI generates a technical implementation plan based on the PRD. This ensures the agent understands the architecture before execution.
  2. Milestone Logs: After each milestone, the AI generates a milestone_log.md. This file serves as a hand-off document for the next phase, ensuring the AI retains context of previous technical decisions.
  3. Self-Verification: The agent is instructed to run automated tests and perform programmatic checks. This 'tier one' QA ensures that the code meets the defined requirements before the human performs manual browser-based testing.

Rapid Refinement

Post-milestone, the developer enters a 'rapid-fire' refinement phase. Instead of dictating specific code changes, the developer describes the problem or the confusion in the UI. This allows the AI to propose better design or interaction solutions. By maintaining the current context, the developer can batch multiple small fixes, keeping the momentum high while ensuring the application remains intuitive and clean.

Key Takeaways

  • Start with the 'Why': Your domain expertise is your greatest asset; define the problem clearly before asking the AI to solve it.
  • Use Plan Mode: Always force the agent to create a technical implementation plan before it starts writing code.
  • Keep a Source of Truth: Store your PRD as an HTML or Markdown file within the repository so it remains a living document.
  • Implement Milestone Logs: Have the AI document its own technical decisions at the end of every milestone to maintain continuity.
  • Define 'Done': Use clear checklists for each milestone to ensure the AI knows exactly what constitutes a completed feature.

Notable Quotes

  • "Whatever you do for a living, you know your work better than any software company or developer ever will."
  • "The hard part was never the code; it's knowing what to build."
  • "I take what the work really needs and turn it into a real tool."
  • "I don't want to do the email invoices, so I'm just going to clarify this one."