用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
直接命令不会经过审查 Prompt；运行前请先检查来源。
npx skills add https://github.com/curiositech/windags-skills --skill thinking-inside-the-box-the-shadowbox-me命令会保持在同一行。复制前请横向滚动并检查完整内容。
想先保存到本地？可下载 SkillsMP 当前能够提供的文件。
正在显示 SKILL.md
基于 SOC 职业分类
| name | thinking-inside-the-box-the-shadowbox-me |
| description | license: Apache-2.0 NOT for unrelated tasks outside this domain. |
| license | Apache-2.0 |
| metadata | {"provenance":{"kind":"legacy-recovered","owners":["some-claude-skills"]}} |
license: Apache-2.0
Load this skill when:
Not appropriate for:
The Principle: Expertise develops not by being told what to think, but by experiencing the gap between your thinking and expert thinking, then discovering why that gap exists.
How it works:
Why it matters: Traditional instruction tells you what experts know. Calibration shows you what you don't see yet. The 18% improvement in firefighter decision-making came from self-discovered gaps, not transmitted knowledge.
Application to agent systems: Agents can log their reasoning at decision points, then compare to expert corpora or ensemble predictions. The divergence signal becomes the training gradient, but for cognitive patterns rather than output accuracy.
The Principle: The one-inch box limitation (≈15 words) is not a formatting choice—it's a cognitive tool that forces prioritization and makes implicit thinking explicit.
How it works:
Why it matters: Experts don't just know more—they attend to different things. Unconstrained responses hide this difference. The constraint reveals what experts consider worth their limited attention.
Application to agent systems: Token budgets, summary constraints, or forced-ranking mechanisms can replicate this forcing function. "If you could only log 3 observations, what would they be?" reveals an agent's priority function.
The Principle: Expertise isn't unidimensional. Different "box types" (Attention, Action Priority, Information, Anticipation, Assessment, Monitoring) reveal different facets of expert cognition.
The six lenses:
Why it matters: A novice might get "what to do" right but fail to anticipate consequences, or notice the right cues but prioritize the wrong actions. Multi-dimensional assessment reveals where in the cognitive pipeline expertise is lacking.
Application to agent systems: Different evaluation dimensions for agent reasoning. An agent might retrieve correct information but fail to anticipate second-order effects. Dimensional diagnosis enables targeted improvement.
The Principle: In complex domains, there is no single "correct" answer—only defensible reasoning. Expert consensus (typically 60-90%, never 100%) teaches that quality of rationale matters more than convergence on solutions.
How it works:
Why it matters: Prepares learners for real-world ambiguity. Experts disagree because situations are genuinely complex. Learning to evaluate reasoning quality under uncertainty is more valuable than memorizing "correct" answers.
Application to agent systems: Ensemble-based training where agents learn from distribution of expert responses, not single labels. Calibration to reasoning patterns rather than output tokens.
The Principle: Information is revealed sequentially, and trainees cannot revise earlier responses. This "no look-back" rule creates consequential learning—you live with your early assessments and see how they play out.
How it works:
Why it matters: Mirrors real decision-making where early choices constrain later options and hindsight isn't available. Teaches trainees to recognize when their initial assessment was flawed and why.
Application to agent systems: Non-revisable reasoning traces create accountability for early-stage reasoning. Agents can't retroactively justify decisions with information that came later, forcing better initial assessment.
IF you need to transfer expertise in domains with delayed or ambiguous feedback
THEN use comparison-based calibration rather than outcome-based reinforcement (outcomes may take months/years; expert reasoning is available immediately)
IF experts are scarce or expensive
THEN capture expert panel responses once, replay many times (one expert panel session can train hundreds of people)
IF learners have widely varying skill levels
THEN use self-paced scenario progression where each learner discovers their own gaps (no classroom synchronization needed)
IF the skill involves tacit knowledge ("I just know")
THEN use constrained response formats that force experts to articulate what they're prioritizing
IF you need measurable progress tracking
THEN measure convergence with expert consensus over time, dimensional by box type (quantifiable expertise development)
IF evaluating agent decision-making quality
THEN assess across multiple cognitive dimensions (Attention, Action, Information, Anticipation, Assessment, Monitoring), not just output correctness
IF agent reasoning seems "off" but outputs are technically correct
THEN compare dimensional reasoning patterns to expert baselines (may be right answer, wrong reasoning)
IF agents need to handle novel situations
THEN train on pattern recognition across dimensions rather than memorization of correct responses
IF building multi-agent systems
THEN use expert consensus models (60-90% agreement) rather than requiring perfect alignment
IF learners are advanced enough to benefit from discovery
THEN provide zero explicit instruction—only scenarios, expert responses, and rationale (let them figure out what they're missing)
IF learners need to understand their own thinking
THEN require commitment before revealing expert responses (creates authentic divergence experience)
IF domain has multiple valid approaches
THEN show minority expert positions with rationale (teaches reasoning quality over conformity)
IF learners seem stuck or defensive
THEN emphasize that expert consensus is never 100% and disagreement is normal (reduces ego threat)
| Reference File | When to Load | Key Content |
|---|---|---|
calibration-driven-learning-for-agent-systems.md | Designing learning systems based on comparison rather than instruction; understanding the 18% performance improvement mechanism | Deep dive into Bloom & Broder's 1950 foundational study, the cognitive science of calibration learning, why divergence detection works, comparison to other training methods, implementation requirements |
constraint-as-cognitive-forcing-function.md | Designing response formats, dealing with verbose or unfocused outputs, making tacit knowledge explicit | The psychology of the one-inch box, why constraint improves thinking quality, how experts use constraints differently than novices, practical guidelines for constraint design, failure modes of unconstrained responses |
expert-consensus-without-ground-truth.md | Handling domains with no clear right answer, dealing with expert disagreement, teaching judgment under uncertainty | How to work with expert panels, consensus thresholds (60-90%), what to do with minority positions, teaching defensible reasoning vs. correct answers, handling learner confusion when experts disagree |
scenario-based-expertise-transfer.md | Creating training scenarios, choosing realistic cases, understanding why abstract principles fail | What makes a good scenario, decision point identification, realism vs. pedagogical clarity, scenario length and complexity, why context-rich cases beat principle-based instruction |
progressive-revelation-and-commitment.md | Designing multi-stage scenarios, creating irrevocable commitment mechanisms, teaching consequence awareness | The "no look-back" rule, why sequential revelation matters, how to structure decision points, commitment mechanisms, how initial framing affects downstream reasoning, retroactive justification problems |
multi-dimensional-cognitive-assessment.md | Evaluating agents across cognitive dimensions, diagnosing specific expertise gaps, understanding different box types | Detailed breakdown of all six box types (Attention, Action Priority, Information, Anticipation, Assessment, Monitoring), what each reveals, how to choose box types for decision points, dimensional diagnosis frameworks |
The Mistake: Adding explanatory lectures, principles, or "here's what you should learn from this" guidance.
Why It Fails: Violates the self-directed discovery principle. When you tell learners what the gap means, you rob them of the insight. The method works because learners figure it out themselves.
What to Do Instead: Trust the comparison process. Provide scenarios, expert responses with rationale, and silence. If learners ask "what should I learn?", redirect: "What differences do you notice?"
The Mistake: Removing scenarios where experts disagree, or trying to adjudicate "who's right."
Why It Fails: Real expertise involves defensible reasoning under uncertainty, not convergence on single answers. Disagreement teaches reasoning quality matters more than conformity.
What to Do Instead: Embrace 60-90% consensus as ideal. Share minority positions with full rationale. Teach learners to evaluate reasoning, not count votes.
The Mistake: Letting learners go back and change earlier boxes after seeing new information or expert responses.
Why It Fails: Eliminates the consequence of early framing decisions. Enables retroactive justification rather than learning from flawed initial assessment.
What to Do Instead: Lock responses once submitted. Make the "no look-back" rule explicit and inviolable. The discomfort of living with early mistakes is where learning happens.
The Mistake: Creating simplified, clean scenarios that lack the ambiguity and complexity of real situations.
Why It Fails: Experts developed their pattern recognition in messy, realistic contexts. Toy problems don't activate the same cognitive patterns and don't transfer to real performance.
What to Do Instead: Use scenarios drawn from actual incidents, with realistic ambiguity, incomplete information, and time pressure. Messiness is a feature, not a bug.
The Mistake: Always using the same box type, or randomly selecting box types without considering what each reveals.
Why It Fails: Different dimensions assess different aspects of expertise. A trainee might excel at noticing (Attention) but fail at prioritizing (Action). Dimensional blindness misses diagnostic opportunities.
What to Do Instead: Strategically select box types for each decision point based on what that moment in the scenario should reveal. Use dimensional patterns over time to diagnose specific expertise gaps.
The Mistake: Evaluating whether trainee responses "match" expert responses rather than whether reasoning patterns align.
Why It Fails: Experts sometimes reach similar conclusions via different reasoning. Novices sometimes guess correctly with poor reasoning. The method targets cognitive patterns, not output tokens.
What to Do Instead: Compare rationale, not just conclusions. A trainee who identifies different cues but with sound reasoning may be developing expertise differently than the panel—explore rather than correct.
They understand calibration vs. instruction:
They grasp the forcing function:
They work with uncertainty:
They respect the discovery process:
They think dimensionally:
They recognize the paradigm shift:
This skill operationalizes expertise transfer through comparison-based calibration. Load reference documents as needed for implementation details.
