What if your AI coding assistant refused to write code until it actually understood what you wanted?
That's the idea behind Ouroboros, an open-source Claude Code plugin that forces a structured interview before any code generation happens. Instead of letting the AI fill in gaps with assumptions, it scores the ambiguity of your request on a 0-to-1 scale. If the score is above 0.2, it keeps asking questions. No code gets written until the spec is locked down.
The scoring formula weights different dimensions depending on your situation. For brand-new projects, it measures goal clarity (40%), constraint clarity (30%), and success criteria (30%). Each dimension gets scored by an LLM with its randomness dialed nearly to zero for consistent results. Once you clear the threshold, the system moves into an execution loop that cycles through interview, seed, execute, and evaluate phases until consecutive outputs are 95% or more similar - meaning the spec has converged and stopped changing.
Under the hood, it's a substantial system: 18 packages across 166 modules, with nine specialized agents handling different parts of the process. These include an interviewer, an architect, an evaluator, and even a "contrarian" agent that challenges assumptions. It supports over 100 LLMs through a routing layer and includes cost optimization that picks cheaper models for simpler tasks.
The project has crossed 1,000 GitHub stars with 34 releases since launch. Installation takes two commands if you already use Claude Code:
claude plugin marketplace add Q00/ouroboros
claude plugin install ouroboros@ouroboros
The core insight is solid: most AI coding failures happen at the requirements stage, not the implementation stage. Anyone who's spent an hour debugging AI-generated code only to realize it solved the wrong problem will recognize the pain point. Whether a formal ambiguity-scoring system is the right solution or just adds friction will depend on your workflow. But at minimum, it's a more disciplined approach than "prompt and pray."