agent-authoring-toolkit

Autonomously improves a Claude Code skill in a review→fix→re-review loop: each round invokes the reviewing-claude-skills review non-interactively, applies every blocking finding — High and structural findings included — without per-fix approval, commits the round, and re-reviews, until the review's verdict is acceptable, the blocking findings plateau, or the round cap is hit. Advisory findings are reported, never chased. Asks one question only: confirming the intent brief at kickoff. Use when the user asks to improve a skill autonomously, in a loop, or until it passes review.

Autonomously improves a Claude Code subagent definition in a review→fix→re-review loop: each round invokes the reviewing-claude-subagents review non-interactively, applies every blocking finding — High and structural findings included — without per-fix approval, commits the round, and re-reviews, until the review's verdict is acceptable, the blocking findings plateau, or the round cap is hit. A gated verdict recommending a different artifact form — 'this should be a skill' — ends the loop for the human instead of being applied. Advisory findings are reported, never chased. Asks one question only: confirming the intent brief at kickoff. Use when the user asks to improve a subagent autonomously, in a loop, or until it passes review.

Reviews a Claude Code skill — its SKILL.md, evals, and referenced files — in two passes. Structure first: a workflow whose shape fails High stops at a structural verdict with a redesign recommendation. Otherwise a full detail sweep against skill-authoring and prompting best practices plus the host project's conventions produces a severity-ranked gap analysis ending in a computed verdict — acceptable, or not yet — honoring the skill's recorded waivers, optionally applying approved fixes. Use when the user asks to review, audit, or improve a skill.

Reviews a Claude Code subagent definition — its frontmatter, body, declared tools, and the siblings it competes with for routing — in two passes. Structure first: a definition whose form or shape fails High stops at a structural verdict with a redesign recommendation. Otherwise a full detail sweep against subagent-authoring and prompting best practices plus the host project's conventions produces a severity-ranked gap analysis ending in a computed verdict — acceptable, or not yet — honoring the subagent's recorded waivers, optionally applying approved fixes. Use when the user asks to review, audit, or improve a subagent or an agent definition.