An ambiguity gate for AI coding agents: measure the request before writing code

AI coding agents often jump into writing code based on a one-line request, leaving critical details unclear, such as data models, storage, priority rules, and ID schemes. This leads to time-consuming revisions and rewrites. The "ambiguity gate" aims to address this issue by measuring the vagueness of the request and preventing the system from freezing a specification until a certain threshold is met.

The Ouroboros project, released under the MIT open-source license, implements this gate with concrete constants. The interview loop assesses clarity across three axes—goal clarity, constraint clarity, and success-criteria clarity—each weighted at 0.40, 0.30, and 0.30, respectively. Ambiguity is calculated as one minus the weighted sum, and a request is considered ready when ambiguity falls below 0.2.

Each axis also has its own minimum threshold (0.75 for goal clarity, 0.65 for constraint clarity, and 0.70 for success-criteria clarity), ensuring that a vague goal cannot be compensated by precise constraints. The gate is a default setting, not a hard stop, and the UI provides an option to generate the specification despite the ambiguity.

The gate's purpose is to make vagueness a deliberate decision rather than an accidental occurrence. It shifts the iteration process from after code generation to the interview phase, where a wrong assumption only requires answering one question instead of discarding the entire implementation. The number assigned to the threshold does not guarantee quality but ensures that the decision to proceed was made intentionally.

To test the gate, clone the Ouroboros repository, run the setup and interview scripts, and provide the starting score along with the first question that uncovers an assumption you hadn't previously considered.