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/m-hosseinpour/agent-skills --skill curating-global-instructions시작 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.
Curating Global Instructions
The global instruction file is read at the start of every session, whatever the project. Its length is a cost paid by every session that follows, so it should hold what governs the work: rules, invariants and decisions that must not be re-derived.
Core principle: every section is either a rule — something that must fire unconditionally — or a technique, which needs a trigger. Rules stay. Techniques move to a skill, where they load when wanted.
The one question
Ask it of every section, and of everything you are about to add:
Is this a rule, or a technique?
A rule — a constraint, a decision, a precedence — fires whether or not anything else loads. It stays. A technique — how to do something, once you know you are doing it — has a trigger, and belongs in a skill. Adding without asking this is how the file grows; trimming without asking it is how a rule disappears.
Before moving anything to a skill
Compare against the skill's BODY, not its description. A description answers "should I open this?"; it does not answer "what is inside?". A section is a duplicate only if the body already carries it — and then it is pure duplication, however useful it reads.
Test the trigger, honestly. Content may move only when the trigger names a situation the agent can recognise from the request: "I am building a UI" and "review this diff" pass. It fails when the trigger requires the skill's own diagnosis first — "the code feels brittle", "cleanup looks missing" — because the agent cannot know it needs the skill until it has already learned what the skill teaches. Where the trigger is weak, keep the rule and point at the skill in one line.
Write the new skill's description as a trigger, not a summary — a summary becomes a shortcut the agent follows instead of reading the body.
When the skill you are drafting is more than a few lines, read writing-skills first if it is available: it covers naming, placement and how to check the result.
What never moves
- A decision no skill can enforce. A skill can be read; it cannot compel.
- A rule that contradicts a harness default — this file is the only place that can override one.
- The single failure that costs most when no skill is opened. State it outright even though a skill covers it; that duplication is deliberate.
Working the file
Measure, do not guess. Count each section before and after. "Feels long" is not a measurement, and a total that moved without a per-section number cannot be reported.
One move at a time, verified. After moving a section, grep the names, paths and terms it carried — an orphaned reference to a section that no longer exists is the failure this work produces.
Replace or delete a superseded line; never banner it. A line explaining what the file used to say is itself a cost paid every session.
Common mistakes
| Mistake | Fix |
|---|---|
| Trimming by feel, or toward a target percentage | Measure per section and report the numbers |
| Deleting a section because a skill "covers that area" | Read the body — areas are not contents |
| Moving content whose trigger needs the skill's own diagnosis (e.g. "the code feels brittle") | The agent cannot recognise the problem until the content has taught it; keep the rule |
| A description that summarizes the new skill's steps | Make it a trigger; see writing-skills |