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/thananon/9arm-skills --skill qwenchance시작 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.
Staying on Track
Long, multi-step work fails three ways: looping, over-thinking, and running out of context. Run the checklist below before each step. When a trigger fires, do the matching action — don't deliberate about it.
Before each step — run this
| Check | Trigger fires when... | Do this |
|---|---|---|
| Looping? | You're about to repeat an action (see signals below) | Break the loop — pick one fix below |
| Over-thinking? | You've reasoned past ~1000 words without acting | Stop. Act on your current best decision, or ask the user one question |
| Context tight? | A low-context reminder appeared, or 2+ budget signals hold | Finish this step, then hand off |
If nothing fires, take the step.
1. Loops — detect and break
A step is a loop if any of these is true:
- You're re-reading a file you already read this session (and it has not changed since).
- You're re-running a command/tool with the same args, expecting the same result.
- You're returning to a hypothesis you already tried and dropped.
- You're "reconsidering from the start" with no new evidence.
- The last 2 steps gained no new information.
Re-reading a file you just edited is NOT a loop — that's verifying.
When a loop fires, stop and do exactly one:
- State the blocker in one sentence and ask the user a specific question.
- Write what you know vs. don't know, then take a different action than last time.
- Looped 2+ times on the same sub-problem? Declare it unsolved-for-now; move on or hand off.
Never repeat a failed action hoping for a different result.
Retry cap: never run the same failing command a 3rd time. Can't get something working (a command, a test runner, an import) after ~3 attempts — even varied ones — STOP and ask the user; don't grind through more variations.
Don't edit blind — it's the top loop source. Read enough to know the change is correct before editing. After each edit, verify it (read the diff / run it / run the test) before the next step. One edit → one check.
2. Thinking — keep it bounded
Cap reasoning at ~1000 words per step. Past that, you're deliberating instead of acting.
- Decide → act → observe. Don't re-derive a decision you already made.
- Can't decide in ~1000 words? The task is underspecified — ask the user one sharp question.
- Don't restate the whole problem to yourself. Reference what you concluded; don't rebuild it.
3. Context budget — count signals, don't estimate
Authoritative: A <system-reminder> about low context / approaching auto-compaction. → Hand off now (section 4). Don't start new work.
Otherwise, count how many of these are true right now:
- [ ] 20+ assistant turns into the task.
- [ ] Read 5+ files, or any one huge file/log/dump.
- [ ] Long tool outputs you keep scrolling back to.
- [ ] 3+ plan steps still left.
Count the boxes that are true, then map the count to an action:
- Count is 0 or 1 → CONTINUE working normally.
- Count is 2, 3, or 4 → HAND OFF — finish the current step, then go to section 4.
Count first, then decide — don't judge by feel. A higher count means more context pressure, not less. Being on the last step or "almost done" does not lower the count or cancel a HAND OFF.
Before any expensive step (large read, new subtask, long generation), ask: "Room to finish this AND hand off after?" If the count says HAND OFF, finish the current atomic unit, then hand off — don't start the next.
4. Hand off cleanly
When context is tight or the user asks:
- Land durable artifacts first — save the file, commit, write the result. Nothing lost.
- Invoke the `handoff` skill to compact the conversation. Don't hand-write the handoff.
- Tell the user plainly: "Context is getting tight — handing off now; start a fresh session (
/clear)."
