Table of Contents
Why It Exists
On 2026-05-22, a real production session was given this condition:
/goal "후속 작업 모두 직접 진행하라"
("Just keep going on all the follow-up work yourself.")
It looked harmless. It triggered an infinite stop-hook cycle.
Here is why:
- "모두 / 끝까지 / fully / every" are universal quantifiers. They sweep in
work Claude is structurally not allowed to do — pulling secrets, opening
third-party accounts, signing legal docs, executing payments.
- Claude finishes the in-scope work, the goal evaluator says
"goal not yet met → continue."
- Claude runs again. Same answer. Same verdict.
- Cycle 1, 2, 3 … cycle 9, where the built-in
stop_hook_active cap finally
cuts the conversation off. Roughly 9× the normal token spend, zero new progress.
This plugin exists so that condition never reaches /goal in the first place.
It interviews you, rewrites quantifiers, demands a measurable success signal,
adds a turn limit, separates session-out items, and prints a footer
listing what Claude is structurally allowed to do.
What It Does
You: /goal-maker "ship the new landing-page A/B test"
Claude: (5–7 quick questions, ~60 seconds)
Result: A finished /goal condition, copied to your clipboard,
saved to ~/Desktop, and previewed in the terminal.
The output is a 4-section prompt with every guard applied:
아래 목적을 진행하여 테스트 성공까지 진행하라. 필요한 정보가 있다면 브레인스토밍하라.
# 목적
- (your one-line objective, with universal quantifiers rewritten)
# 조건 및 상황
- (3–5 constraints you supplied)
# 필수 결과물
- (the deliverable that means "done")
## 세부사항
- **측정 기준**: (binary, machine-checkable signal)
- **증명 채널** (Claude가 transcript에 surface할 형식):
- (exactly how Claude must print the result so the evaluator can see it)
- **턴 제한**: 최대 20턴 안에 condition 충족 또는 명시적 사용자 승인 요청
- (your end-of-workflow step, e.g. /handoff-verify → /commit-push-pr --merge)
- 테스트 통과를 위한 테스트 삭제 금지, 기존 테스트 카운트 유지
- 단, 본 작업 외 무관 리팩토링 / 의존성 업그레이드 / 다른 도메인 변경 금지
### 세션 외 검증 (Plan에 명시 후 사용자 위임)
- (external dependencies pulled out so the evaluator stops blocking on them)
## 세부 목적
- 궁극적으로 (your business outcome) 좋아져야 한다.
---
_session 한계: 시크릿 추출 / 외부 계정 신규 발급 / 사용자 명시 승인이 필요한 항목은 Claude 권한 밖이며, 해당 항목 발견 시 Plan에 명시 후 사용자 안내._
The 8 Guards
| # | Guard | What it prevents |
|---|
| G1 | Universal-quantifier rewrite | "모두 / 끝까지 / every / fully" silently sweeping in out-of-scope work |
| G2 | Measurable signal | "잘 동작" / "perfect" — vague success that the evaluator can never confirm |
| G3 | Scope guard | Unrequested refactors, dep bumps, drive-by edits in unrelated files |
| G4 | Letter-vs-spirit | "Passing tests" by deleting tests or xit-ing them out |
| G5 | Session-limit footer | Treating Claude as if it could pull secrets or open Stripe accounts |
| G6 | Proof channel | Measurements that never surface in the transcript → evaluator can't judge |
| G7 | Turn limit (default 20) | Runaway loops outliving the built-in stop_hook_active cap |
| G8 | Session-out separation | "User must demo it / GSC must index it" blocking the goal forever |
G1–G5 came from the 2026-05-22 incident.
G6–G8 came from Anthropic's official /goal guidance for evaluator-friendly conditions.
Install
Option A — Via the QJC plugin marketplace (recommended)
