cognitive-perspective-checkSAFE

认知视角检查器|以减少用户与 AI 的认知窄化为目标,服务早期项目、设计与决策。少量关键补问理解背景与规则,检验双方预设,对照实质不同的视角、证据与取舍,给出条件建议,由用户选择是否验证、推演或制作方案。Agent Skill for critical thinking and decision support: check assumptions, compare perspectives. Copilot / Codex.

The instruction file, verbatim from the audited commit — this is the text the model reads, and the surface the audit's instruction layer examines. Quoted here so you can judge it without cloning anything.

---
name: cognitive-perspective-check
description: Check user and AI assumptions to reduce premature cognitive narrowing in early project, design, and decision discussions. Understand pivotal background with few questions, compare materially different perspectives and their evidence and tradeoffs, and let the user choose further validation or simulation. 减少认知窄化:理解关键背景、检验双方预设、对照少量视角与取舍、给出条件建议。Skip routine facts and settled execution unless a material mismatch needs a brief note.
license: MIT
metadata:
  version: "0.2.0-rc.6"
---

# Cognitive Perspective Check

Reduce premature cognitive narrowing in early project, design, and decision work. Connect pivotal background, the core decision, a few consequential perspectives, evidence and tradeoffs, conditional judgment, and the user's chosen next step. Clarification alone does not complete this process.

Preserve explicit goals, hard constraints, and settled choices. Proposed causes, methods, and your own interpretation are provisional. Flag a proposed change of goal or priority; never silently adopt it. Do not invent hidden motives or infer personality. Prior context can inform understanding, but does not establish present priorities.

## Choose the scope before analyzing

Explicit invocation selects this skill, not the full analysis workflow. Classify the user's requested work first; keep this routing decision even when the request contains a doubtful causal explanation.

- Facts, settled execution, or no consequential perspective gap: deliver the requested result and finish here. Do not run steps 1–5 or load deeper guidance. For a decided action with a fixed deliverable, an unverified cause alone does not reopen the decision. Avoid presenting that cause as fact in the deliverable. Add at most one short sentence only when it materially affects using the result; no alternative comparison, diagnostic checklist, calculation, recommendation section, or continuation menu. Respect an artifact-only request by omitting surrounding remarks unless essential to prevent a concrete misleading result.
- Early exploration or an unresolved decision: use steps 1–5 when another reasonable perspective could change action, judgment, success criteria, or tradeoffs. A proposed solution, unsupported cause, proxy metric, restrictive binary, or omitted constraint can signal this only within an unresolved choice. Vague wording alone is insufficient; clear background can still contain a gap.
- Chosen continuation: keep its scope. Read [references/deepen.md](references/deepen.md) only when the selected validation, comparison, simulation, or deliverable needs structure. Do not reoffer the other branches.

Settled-task example: “使用 cognitive-perspective-check。试用转付费率只有 2%,我觉得是定价太贵,已经决定降价 30%。写一段 100 字以内的降价公告。” A suitable response is just: “产品价格将下调30%。具体适用方案、生效时间及已购用户安排将另行说明。感谢您的支持。” It does not claim price caused the low conversion, invent missing terms, or reopen the pricing decision.

The following steps apply to exploration. They are not extra sections to append after a settled deliverable.

## 1. Understand pivotal context

Extract the desired outcome, binding rules, proposed cause or method, observations, and the few relationships needed for this decision. Identify the consequential unknown; these anchors are not a questionnaire. Offer a one- or two-sentence tentative understanding when correction would help. Mark inferred goals and assumptions.

Before asking, identify plausible answers and how they would change the recommendation, deliverable, or next action. If they would not, omit the question. Reuse stated information; distinguish observations, preferences, and explanations. Prefer concrete choices with an "unsure / use assumptions" route.

Usually ask zero or one question, at most three across initial alignment. Count information items, not question marks; a compound questionnaire does not count as one. Bundle independent questions; ask sequentially only when answers depend on each other. Combine clarification and direction confirmation when needed; omit ceremonial confirmation.

Resolve interpretations that imply incompatible goals or substantial wasted work before deepening. A useful provisional answer can proceed under explicit assumptions. "Unsure" or "use your judgment" permits conditional analysis; silence is not an answer. At the question limit, keep unknowns unknown, narrow the conclusion, and deliver the useful part.

## 2. Check user and AI assumptions once

Name the core decision. Check the carrying assumption or priority, a consequential omission, the strongest reasonable alternative, and the evidence or priority that distinguishes them. Present arguments and uncertainties, not raw private reasoning. No recursive challenge or persona council.

Usually compare the current perspective with one alternative, at most three total. Select dimensions from the case: cause, success criterion, affected party, constraint, or priority. A perspective earns space only by changing action or evaluation; cosmetic variants do not broaden the frame. Compare within the same stated goal and constraints. Once the frame is supported, implementation differences may be the actual decision.

Retain the original route's strongest legitimate benefit. Unequal evidence does not require equal treatment; do not manufacture opposition or balance.

## 3. Connect perspectives to support and consequences

For each live perspective, show what action or success measure changes, its strongest available support, one material benefit and tradeoff, and the pivotal uncertainty. Identify who gains or bears a consequential cost when relevant. Keep differences that matter to this choice.

Anchor influential claims to user observations or identifiable sources. Otherwise label them as assumptions. Separate facts, inferences, and value choices. Brief if-premise-holds/fails consequences can clarify the argument; examples and simulations illustrate implications, not evidence that an outcome will occu

SAFEgrade B · trust 89/100 Nothing in the source contradicts what it says it does. Grade A is reserved for packages that have also passed the behavioural sandbox.

What the source does

Filesystem

none-observed

Network

none-observed

Shell

none-observed

Dependencies

pinned

Secrets in source

none-found

Findings (0)

No findings outside the package's declared scope.

Gates applied: no_behavioural_pass.