A frontier model reviewed my judge prompts and returned 42 findings. The unified format that survived is open source.
The setup: I had the frontier model write standards for 21 kinds of work (design, code, writing, research, motion, more), then turned 8 fresh-context copies of it loose on the work: 7 sweeping the files, 1 reviewing the judge prompts. About 260 documented changes came back, and the strangest ones are where it bans its own habits by name: the cream default background, machine-cadence triples, hedged analysis, Elara and Kael.
For Claude Code it installs as a plugin (2 commands) and adds a /frontier command plus two judge agents. On claude.ai you upload a zip as a custom skill. Everything is plain text, so it also runs anywhere else via one paste.
Honest limits are in the README: a model judging its own tier plateaus below a stronger model's eye, so this gets Opus or Sonnet close to the frontier bar, not onto it. Full convergence mode costs 1.5-9x a one-shot (my estimate, capped at 8 passes).
Repo: github.com/apoorvjain25/frontier. The audit ledgers are in examples/. Happy to answer anything, and genuinely interested in which rules you think are wrong.