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/terrylica/cc-skills --skill crucible-navigator시작 전 확인
- 설치에는 모델 구독, API 한도, 파일 접근 권한이 포함되지 않습니다.
- 별점 수와 설치 수는 활동 지표이며 독립적인 품질·보안 검증이 아닙니다.
원본 Skill 문서
원본 저장소의 제목, 예시, 코드, 표, 링크, 이미지를 유지해 표시합니다.
crucible-navigator — Research campaign orientation
Self-Evolving Skill: This skill improves through use. If the routing guidance is wrong (wrong skill invoked for a user intent), fix the description's TRIGGERS list and the routing table below. Append reason to references/evolution-log.md. Don't defer.What this skill does
You are orienting an agent to research methodology. Route the agent to the right sub-skill based on what kind of research task is at hand. The four sub-skills cover epistemics (A), execution (B), decisions (C), and archive (D).
Routing table — map user intent to skill
| User intent signal | Route to |
|---|---|
| "Is this finding real?", "check significance", "shuffled null", "causal feature", "label leakage" | a-research-foundations |
| "Test a hypothesis", "sweep parameters", "multi-agent analysis", "gate validation", "per-trade analysis" | b-investigation-methodology |
| "Should we pivot?", "ship this?", "kill or refine?", "too much context", "supersede a finding" | c-meta-governance |
| "Check dormant ideas", "what's archived?", "resurrect this", "conditions have changed" | d-emergent-resurrection |
If unsure, start with A (foundations). Epistemic discipline is the hardest prerequisite.
The core genetic-evolutionary framing
Research in this plugin is treated as a population under selection:
Population ───────────> (multi-lens agents propose candidates)
│
▼
Fitness ──────────────> (z-score vs shuffled null, OOS replication, kill-selectivity)
│
▼
Selection ────────────> (serial adversarial gates A→B→C→D→E)
│
▼
Surviving individuals ─> (champion stacked as supersedes chain in evolution.jsonl)
│
▼
Mutation/Crossover ───> (agent proposes perturbations of winner; orthogonal components recombined)
│
▼
Archive failures ─────> (d-emergent-resurrection preserves with resurrect_if conditions)
│
└──────── NEXT GENERATION ───────┐
▼
(new audit folder)A "hypothesis" is a triple: (trigger, filter_cascade, management_rule). Each component mutates independently. Crossover combines validated components from different campaigns. Failed individuals go to archive, not trash.
Reading order for a new campaign
- Read this file. Understand the genetic framing.
- Read `a-research-foundations/SKILL.md`. 6 epistemic disciplines.
- Read `b-investigation-methodology/SKILL.md`. 6 execution patterns. This is the workhorse.
- Reference `c-meta-governance/SKILL.md` as needed during pivots/decisions.
- Check `d-emergent-resurrection/SKILL.md` first if the hypothesis resembles an archived dead one — you might be resurrecting.
Anti-patterns
- Treating principles as universal from N=1 evidence. Each principle has a
confirmation_count; defer to re-confirmed ones. - Skipping A. Experienced agents skip epistemics — don't. The label-leakage trap, wrong-null trap, and agent-z overstatement trap all live in A.
- Running a new campaign without checking D. Ideas recur. Check archive before re-exploring.
- Writing conclusions without ledger entries. See A5 (record-keeping-discipline).
Repository landmarks
In research-lab-repo/:
findings/methodology/— original 10 principle files (pre-plugin)findings/evolution/evolution.jsonl— ledger of findings (append-only)findings/evolution/audits/— dated audit folders (one per campaign)findings/evolution/audits/CLAUDE.md— audit index
When you're done in a session
Update references/evolution-log.md if you noticed that:
- The routing table missed a user intent (add the row)
- A TRIGGER keyword didn't fire on a relevant prompt (extend the description)
- A sub-skill was weaker than expected (flag it for refinement)
Post-Execution Reflection
After invoking this skill, if the routing produced a wrong-skill-selected outcome, or if a new user-intent pattern emerged that isn't covered:
- Identify the gap. What did the user ask for that wasn't routed?
- Determine the fix. Add a new row to the routing table? Add keywords to TRIGGERS? New sub-skill needed?
- Apply the fix. Edit this file in-place. Append to
references/evolution-log.mdwith: date, trigger text, what was changed, and link to the session where this came up. - Never silently defer. A quiet gap is worse than a noisy one.
