Tools
Things I have built that come out of the research. Each one is public, and each one carries the evidence it rests on. Nothing here is a product, and none of it is supported.
-
A gate that refuses to let an AI agent start on an underspecified request, and a committed record of what was missing.
Coding agents begin as soon as they are asked, whatever the request leaves out, and what they leave out is not random: security, compliance and operational requirements are the least covered. spec-gate scores a request against a profile, refuses below a threshold, and writes a record of what was missing, what was assumed, and who authorised proceeding. Deterministic rules cap what the agent may claim, so it cannot declare its way through, and a hook stops edits that no record authorises. That record is the point. It is what a reviewer can read months later, and what a chat transcript can never be.
Across five models on 100 underspecified requests, gating cut critical omissions by 29 to 50 percent. Clarifying questions without a stopping rule made them worse on four of the five.
Working, unreleased Repository →The server, the gate, the records and the enforcement hook all work, under 47 tests. Not published to npm, and not yet proven in real use.
Two conditions for anything on this page: it is public and installable, and it carries evidence you can check. A tool with neither belongs in a drawer, not on a site.