A developer tested Claude by asking it to "make my scraper robust," receiving 200 lines of plausible-looking code. The output was fundamentally flawed: retry logic mismatched the codebase, global logging broke modules, and pagination lacked a guard against infinite loops. The code looked professional but was built on assumptions, not understanding.

The core problem with letting AI code without planning is that it guesses patterns, confirms its own blind spots, and ships the first idea too quickly. By the time developers catch the issues, they're already committed to a bad approach, making rework expensive. This reveals a critical gap in current AI coding workflows.

Instead, the developer forces Claude to argue with itself before writing any code. A structured conversation involves multiple personas: a planner proposes an approach, a critic challenges assumptions, a builder writes the code, and a reviewer validates it. This "plan, critique, build, validate" pattern surfaces bugs before they're coded.

This method mirrors emerging AI coding tools like Devin and Cursor's agent mode, which are converging on structured planning. The developer has open-sourced these Claude skills as a team, allowing anyone to clone and use the multi-voice approach for more reliable AI-generated code.