Builds a project's AI rule set: a slim CLAUDE.md index, an AGENTS.md pointer for Codex, and on-demand docs/*.md rule files.
npx skills add moveju112/rule-architect -gWhy
One CLAUDE.md that grows to a few hundred lines gets loaded into every session whether or not the rules apply. This skill splits it: a short always-loaded index plus topic files the agent reads only when relevant.
What's in this release
Bilingual. Triggers and output work in Korean and English. Structural tokens stay fixed in every language — MUST / NEVER / PREFER, why:, the ❌/✅ markers, uppercase file names, and the generation marker.
Installable three ways. Skills CLI (npx skills add), Claude Code plugin (/plugin marketplace add moveju112/rule-architect), or a plain git clone. The plugin path is new in this release — the manifest previously sat at the repo root, where Claude Code does not look for it.
Verified output. Generated rule sets are checked for broken links and token budget, then quizzed — the run answers questions against its own rules to catch ones that read clearly but do not actually decide anything.
--update does a diff-update. Only changed sections are rewritten, so hand-added rules survive a refresh.
Before you run --update
This skill leans on judgment. A weak model regenerates the whole file instead of diffing and will erase rules you added by hand. Use Opus 5 / Fable 5 with high effort on Claude, or a flagship GPT-5 tier on Codex — never a mini or nano class. Commit your existing CLAUDE.md first.