frontier
A written set of quality standards and a convergence procedure that any language model can run itself through, so its output is judged against explicit rubric lines and fresh-eyes verification instead of a vague request to make it good.
frontier is a Claude Code command that runs a task through a written quality procedure rather than asking a model to simply try harder. The premise is that the gap between an average model response and a strong one is mostly method, not raw capability: define the standard before generating, sample several attempts instead of polishing the first one, verify against real evidence with a fresh perspective, and refuse to stop at something that merely looks finished.
It ships 21 craft standards, and offers a single-response mode where a model drafts against a standard, then runs one fresh-eyes judgment pass on its own output and fixes what that pass finds, all inside one reply; a fuller convergence loop and a final taste-judgment gate are available for work that has to be right. Findings are reported in a fixed, parseable shape, pinned to a location and a specific rubric line, rather than as unstructured commentary, and an empty findings list is treated as a claim that every rubric line was actively checked.