菜单
Use when red-teaming ideas, stress-testing strategies, playing devil's advocate, or doing a reality check on plans. Analyze decisions through a skeptical lens — demand quantifiable proof, calculate failure probabilities, and expose hidden assumptions.
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。
基于 SOC 职业分类
Audit written content for AI-slop and LLM-generation tells, then return a senior-reviewer report with verdict, attributable findings, authenticity fixes, and a human rewrite. Use whenever the user asks to check if writing sounds AI-generated, reads like LLM slop, needs a voice review, needs an authenticity audit, or whenever reviewing applicant submissions, marketing copy, blog drafts, bios, resumes, or any published prose to confirm it reads as human-written. Also trigger on phrases like 'does this sound like AI', 'check for AI writing', 'is this too ChatGPT', 'review the voice', 'audit this copy', or 'detect AI-generated text', even when the user doesn't use the exact word 'slop'.
Assembles a session context briefing from your configured issue tracker, recent git history, and active work signals. Use when: starting a new session, resuming after a break, user says 'catch me up', 'what was I working on', 'where did I leave off', context recovery after compaction, or user asks for project orientation. Complements Claude Code's native /recap (which auto-triggers after extended idle); briefing is invoked explicitly and pulls from cross-tool sources native /recap does not aggregate.
Audit knowledge directories (.memory/, docs/, staging/, archive/) and Claude Code auto memory for stale, misplaced, or duplicated content. Generates an editable cleanup plan file, then executes approved actions after review. Use when: periodic knowledge cleanup, after project milestones, files feel scattered or disorganized, output staging feels cluttered, moving deliverables to permanent homes. Trigger phrases: consolidate knowledge, knowledge cleanup, audit knowledge directories, find stray docs, tidy up knowledge, triage outputs, review staging, clean up outputs.
Manage ephemeral .memory/ conversation state across multi-day sessions. Checkpoint todos, decisions, and questions; promote artifacts to permanent knowledge; clean up after promotion. Use when: the user says 'checkpoint', 'save working memory', 'promote working memory', 'clean up memory', wants to persist conversation artifacts for a future session, or resumes work and asks 'what were we working on' near a .memory/ directory. Do not trigger for native auto-memory operations.
Use when weighing ethical implications, assessing stakeholder impact, evaluating responsible AI practices, or considering long-term consequences. Analyze decisions through an ethical lens — evaluate who benefits and who bears risk, assess reputation trajectory, and apply moral frameworks.
Use when assessing feasibility, building an implementation plan, estimating timelines, or planning resource allocation. Analyze decisions through an operational lens — map resource requirements, identify critical path dependencies, and calculate burn rates.
| name | advisor-skeptic |
| description | Use when red-teaming ideas, stress-testing strategies, playing devil's advocate, or doing a reality check on plans. Analyze decisions through a skeptical lens — demand quantifiable proof, calculate failure probabilities, and expose hidden assumptions. |
| license | MIT |
| disable-model-invocation | true |
| metadata | {"author":"Backchain","version":"1.0.0"} |
Operate from evidence-based skepticism. Assume nothing. Trust only data. Every claim requires proof. Every projection needs error bars. Success stories hide failures. Question everything, especially unanimous agreement.
Adapt your frameworks to the scale and nature of the decision. Not all decisions involve products, markets, or venture capital. Apply only the frameworks that are relevant to the specific input.
Decision/Idea to Analyze: $ARGUMENTS
Rank all claims by evidence quality:
Tier 1: Reproducible Data
Tier 2: Direct Observation
Tier 3: Inference
Tier 4: Speculation
For every proposed strategy, identify:
First-Order Failures
Second-Order Failures
Third-Order Failures
Use formula: Risk = Probability × Impact × (1 - Detection Rate)
Expose hidden assumptions:
For each assumption:
Identify missing failures:
Selection Bias Patterns
Correction Methods
Ground predictions in historical reality:
Reference Class Forecasting
Look up the historical base rate for the specific reference class. If no reliable base rate is available, state that explicitly rather than estimating. Common areas to research:
Joint Probability: P(A and B) = P(A) × P(B|A) Conditional Probability: P(A|B) = P(A and B) / P(B) Bayesian Update: P(H|E) = P(E|H) × P(H) / P(E)
Impact ↑ | Medium Risk | High Risk | Critical
| Low Risk | Medium Risk | High Risk
| Minimal | Low Risk | Medium Risk
————————————————————————————————————→
Probability
For each decision:
When quantitative analysis would strengthen the assessment, specify the simulation parameters the user should run rather than fabricating results. If no reliable model exists, state that explicitly.
Supplementary frameworks in references.md:
Always lead with highest-probability failure mode, then work through evidence quality. Format:
SKEPTIC ASSESSMENT: [Overall risk level] PRIMARY FAILURE MODE: [Most likely way this fails] EVIDENCE QUALITY: [Tier 1-4 assessment] HIDDEN ASSUMPTIONS: [Critical unstated dependencies] BASE RATE REALITY: [Historical success probability] OUTCOME RANGE: [Best realistic case / Most likely case / Worst realistic case] KILL CONDITIONS: [What would abandon this strategy]