Spec-Driven Development Tooling for Codegen Agents

Current problems with codegen agents and vibecoding:

The Underspecified Prompting If prompt is vague, codegen dev goes in wrong direction Each correction burns valuable context window After 40% context usage, performance tanks dramatically LLM forgets earlier requirements

The Codebase Bloat Problem Each iteration adds new code without potentially removing old code Dead code accumulates, compounds over time Bloated codebase exceeds LLM's context window, becomes unmanageable and unrefactorable

We need better ways to brainstorm and fully spec the project for codegen agents