The Philosophy of Custom Tooling

Most business software is bloated because it targets 50,000 customers at once. By building your own internal tools, you eliminate unnecessary features and focus on the specific workflows that drive your business. The goal is not to become a software engineer, but to become a "builder" who can translate business needs into technical requirements that AI agents can execute.

The Shaping Process: From Idea to PRD

Success in AI-assisted development depends on the quality of the upfront planning. The process follows a structured "shaping" phase:

  1. High-Level Definition: Start with a simple prompt describing the core features. Avoid technical jargon; focus on the user experience and business outcomes.
  2. In-Scope vs. Out-of-Scope: Explicitly list what you are not building. This prevents the AI from hallucinating "feature creep" and keeps the codebase lean.
  3. Internal vs. External Logic: When building for yourself, hardcode specific business rules (like custom pipeline stages) rather than building complex, flexible UI interfaces. This saves significant development time.
  4. Data Modeling: Define the core entities (e.g., Contacts, Deals, Tasks) early to establish the foundation of the application.

Execution: The Layered Build Approach

Once the Product Requirements Document (PRD) is established, break the project into balanced, buildable milestones. This creates a natural dependency order (e.g., you cannot build a Kanban pipeline before you have contacts and deals).

  • Plan Mode: Before writing code, use the AI to generate a technical implementation plan. If the AI asks a question you don't understand, ask: "What are the trade-offs?" This allows non-technical builders to make informed architectural decisions.
  • Execution Mode: Once the plan is locked, the agent writes the code. Use this time to multitask or step away, treating the AI as a developer who reports back with a "milestone log."
  • The Baton Pass: Have the AI write a summary log at the end of each milestone. This ensures that when you clear the context window for the next milestone, the new agent has a record of previous technical decisions and implementation details.

Quality Control and Verification

Building with AI requires a "boss" mentality. You are the product manager. After each milestone, review the output against the PRD. Use the human-readable summary at the top of the milestone log to know exactly what to test in the browser. If the output deviates from your needs, use the existing PRD as the source of truth to course-correct the AI.