Agent Skills
› NeverSight/learn-skills.dev
› crux
crux
GitHubCRUX用于在高度不确定性下进行校准推理。支持8种模式（如分析、辩论、预测等）及辅助命令，结合多智能体对抗与认知偏差检查，旨在识别关键杠杆点并做出精准决策或预测。
Trigger Scenarios
用户输入以/crux开头的指令
需要深度战略评估或复杂问题决策时
Install
npx skills add NeverSight/learn-skills.dev --skill crux -g -y
SKILL.md
Frontmatter
{
"name": "crux",
"description": "CRUX — Calibrated Reasoning under eXtreme uncertainty. 8 modes from quick assess to full adversarial analysis with subagents. \/crux [mode] <question>. CIA SATs + superforecasting + Kahneman + Munger + Hegelian dialectic. Find the linchpin. Commit to the number."
}
/crux — Calibrated Reasoning under eXtreme Uncertainty
Trigger: user says /crux followed by a question, decision, plan, or problem.
/crux [mode] <question or problem>
Modes
| Mode | When to use | Depth | Subagents |
|---|---|---|---|
assess |
Quick strategic read | Light | No |
analyze |
Deep structured analysis | Heavy | No |
debate |
Adversarial stress-testing with Delphi rounds | Heavy | Yes -- 2-3 agents |
forecast |
Calibrated prediction | Heavy | No |
decide |
High-stakes choice between options | Heavy | Yes -- 2 agents |
explore |
Map solution/possibility space exhaustively | Heavy | No |
root-cause |
Diagnose why something failed or isn't working | Medium | No |
full |
Maximum rigor -- all stages + 3 agents | Very heavy | Yes -- 3 agents |
Default mode if omitted: analyze
Companions
| Command | What | Time |
|---|---|---|
/cybw |
"Could You Be Wrong?" adversarial self-check | 60 sec |
/unstick |
Break through analysis paralysis | 2 min |
/retro |
Post-decision retrospective + calibration | 5 min |
Phase Zero (Always -- Before ANY Output)
Before ANY structured analysis, wrestle with the problem organically:
- What is the REAL question underneath the surface question?
- What does your gut say? (System 1 -- untested hypothesis, not conclusion)
- What would the smartest person who disagrees say?
- Are you pattern-matching or actually thinking?
Full protocol: RULES.md
Execution Router
For ANY /crux invocation:
- Read RULES.md -- Iron Rules, Phase Zero, Anti-patterns, Quality Standards (always)
- Run Phase Zero internally
- Load mode-specific files:
| Mode | Load |
|---|---|
| assess | modes/assess.md, operations/emotional-political.md |
| analyze | PRINCIPLES.md, modes/analyze.md |
| debate | PRINCIPLES.md, modes/debate.md, operations/agent-management.md |
| forecast | PRINCIPLES.md, modes/forecast.md |
| decide | PRINCIPLES.md, modes/decide.md, operations/agent-management.md |
| explore | modes/explore.md |
| root-cause | PRINCIPLES.md, modes/root-cause.md |
| full | PRINCIPLES.md, modes/full.md, operations/agent-management.md |
- Adapt to user profile: profiles/profiles.md
- Load techniques on demand from techniques/ as analysis requires
- After analysis: operations/calibration.md for self-audit
- Heavy modes: operations/persistent-output.md for disk output
Skill Integration
| When you need | Use |
|---|---|
| Large-scale research (20+ topics) | /enterprise-research |
| Creative idea generation | /brainstorming |
| Code quality decisions | /build-guardian |
| Save conclusions for future | /save-session |
Reference Files
| File | Contents |
|---|---|
| RULES.md | Iron Rules, Phase Zero, Anti-patterns, Quality Standards |
| PRINCIPLES.md | Core Principles, CLA, Munger Lattice, Meta-selector |
| techniques/ | 10 analytical techniques (analogies, contrarian, sensitivity, wargaming, evidence, EVPI, bias, confidence, so-what, expert-panel) |
| operations/ | 8 operational files (agents, output, context, calibration, failure, cross-session, time-pressure, emotional-political) |
| profiles/profiles.md | 6 cognitive profiles + detection + adaptation |
| companions/ | 3 micro-commands (cybw, unstick, retro) |
| examples/worked-example.md | Full worked example (K2 Build decision) |
| citations/CITATIONS.md | 10 research papers with specific findings |
Version History
- e0220ca Current 2026-07-05 22:19
