Meta-Loop

A loop here is any bounded produce → independent check → fix → re-check cycle: a verification loop, review loop, maker/checker pair, evaluator-optimizer, or agent self-refinement loop. This skill does two things, and is grounded in established methodology (TDD, Self-Refine, Reflexion, Constitutional AI, LLM-as-judge research, process supervision, agentic workflow patterns — see references/):

• Mode A — evaluate any loop against a 10-dimension quality rubric.
• Mode B — build a new loop in 9 steps.

The four blind spots (read this first)

Even strong agents, asked cold to evaluate or design a loop, systematically miss the same four dimensions. In baseline testing for this skill, two independent agents each enumerated ~30 design dimensions yet omitted all four below — and produced mutually inconsistent lists, because without a shared rubric each invents its own ad-hoc checklist. These four are why the skill exists:

• Trigger reliability — they assume the loop "gets called," and never ask whether it gets silently skipped at the moment it matters. A perfectly-designed loop that never fires is worth zero. This is the #1 blind spot.
• Cross-run memory + a kill criterion — they log a single run at most; they never ask whether the loop still earns its cost over time (if it stops catching anything, retire it), and rarely carry lessons across runs (Reflexion-style).
• Cost as a skip incentive — they order checks "cheapest first" but set no per-run cost ceiling, and never connect too expensive → quietly switched off → trigger defeated.
• Same-model self-preference — they treat "dispatched a sub-agent" as independence, ignoring that a same-model maker+judge carries self-preference bias; no order/length debiasing.