사용 방법
- 설치 명령을 복사하고 자신의 에이전트 환경에 설치하기 전에 소스 패키지를 확인하세요.
- 필요한 도구, 인증 정보, 지원 환경은 원본 문서에서 확인하세요.
- 입력과 원하는 산출물을 설명하고 출처와 결과를 대조하세요.
npx skills add https://github.com/sociuu/sociuu-skills --skill sociuu-better-agent시작 전 확인
- 설치에는 모델 구독, API 한도, 파일 접근 권한이 포함되지 않습니다.
- 별점 수와 설치 수는 활동 지표이며 독립적인 품질·보안 검증이 아닙니다.
원본 저장소의 제목, 예시, 코드, 표, 링크, 이미지를 유지해 표시합니다.
Refactor Agent Instructions
Make repository guidance easier for an agent to load just in time. Preserve the instructions that affect decisions; do not turn a documentation cleanup into a change to product behavior, delivery workflow, or personal settings.
This skill is for refactoring agent instructions, not for implementing a product task. Keep existing repository-specific authority and nested AGENTS.md files intact unless the user explicitly includes them in scope.
Discover before changing
Read the root AGENTS.md, applicable nested AGENTS.md files, and the repository's instruction/source-of-truth documentation. Inventory each rule with its source, intended scope, and whether it is actionable and current. Treat higher-authority project instructions as authoritative rather than as a contradiction with a lower-level convenience note.
Identify genuine conflicts: rules that apply to the same situation but require incompatible behavior, stale instructions that contradict current documented workflow, and duplicated rules whose wording changes the decision. Present each conflict with both sources, its practical consequence, and a recommended resolution. Ask the user which rule to keep before moving, deleting, or rewriting either rule. Do not silently choose between competing policies.
Design the instruction tree
Keep the root AGENTS.md limited to information needed on nearly every task:
- a one-sentence repository description;
- its non-default package manager and non-obvious build, typecheck, or test
entry points; and
- durable scope, safety, or navigation rules that genuinely apply everywhere.
Use the repository's established instructions location. For a new Sociuu-style layout, use .agents/instructions/; do not create a competing docs/, documentation/, specs/, or openspec/ root merely to hold agent rules. If project documentation already owns a documented hierarchy, link to that hierarchy instead. Keep module-specific guidance beside the module or in a nested AGENTS.md when it should not affect other packages.
Group the remaining rules by decision area, not by arbitrary file size. Common groups include workflow and safety, architecture, data/API contracts, testing and verification, security, and documentation. Create only groups that have real, distinct guidance. Each instruction file should say when it applies and link to deeper canonical documents rather than duplicating them.
Propose, then apply
Before editing, show the user:
- unresolved contradictions and the choice needed for each;
- the minimal root content and proposed instruction tree;
- a mapping from every retained rule to its new owner; and
- deletion candidates, labeled redundant, vague, obvious, or
stale, with a concrete reason.
After conflicts are resolved, write the minimal root file, the scoped instruction files, and only the necessary documentation indexes. Use relative Markdown links. Preserve the meaning, authority, and discoverability of retained rules; do not duplicate a rule across root and child files merely for convenience.
Do not delete source instructions, historical records, or an existing instruction tree without explicit authorization. When removal is authorized, prefer a tracked deletion-candidates record when the repository already uses one; otherwise report the candidates in the handoff.
Verify
Read the resulting instruction chain as an agent would: root first, then the files relevant to one representative task from each affected area. Confirm that links resolve, every retained rule has one clear owner, nested guidance is reachable, and the root contains no task-specific or language-specific noise. Run the repository's documentation or catalog validation when available.
