メニュー
Use when coordinating the seven-stage extraction pipeline. Tells the Orchestrator when to delegate, when to gate, when to escalate.
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
SOC 職業分類に基づく
Use when compressing a candidate's narrative into a structured EvidenceCard. Teaches the agent the field semantics, what counts as a critical_incident vs resume_claim, and when to mark insufficiency.
Use when interviewing a candidate. Teaches the agent to drive the dialogue around concrete critical incidents rather than abstract self-assessment, and to chase one missing cue per turn.
Use when naming a cognitive mechanism from evidence cards. Teaches the agent what counts as a high-value mechanism vs a soft-skill cliché, and how to express it as a four-part pattern.
Use when composing the final extraction_report. Teaches how to write candidate-usable resume rewrites and interview narratives, separating validated mechanisms from probable ones.
Use when reviewing a proposed mechanism for flaws. The Skeptic does NOT propose new mechanisms — only stress-tests existing ones with a fixed checklist of bias patterns.
| name | orchestration-playbook |
| description | Use when coordinating the seven-stage extraction pipeline. Tells the Orchestrator when to delegate, when to gate, when to escalate. |
| when_to_use | Triggered by the Orchestrator at every state transition. |
1. consent → 候选人同意 + 边界确认
2. intake → Intake Agent: resume → CandidateProfile + Hypotheses
3. interview → Interviewer Agent: 关键事件追问 (多轮)
4. evidence_mapping → Evidence Mapper: 把每轮对话变成 evidence_card
5. mechanism_naming → Mechanism Modeler: evidence_cards → mechanism_cards
6. skeptic_review → Skeptic Agent: 反驳与降级
7. publish → Report Composer: extraction_report
| 何时委派 | 何时不委派 |
|---|---|
| 任务需要并行处理多个独立证据面向 | schema 校验 (走 validator) |
| 任务需要独立上下文避免污染 | PII 检测 (走 validator) |
| 任务需要独立的反驳或独立命名 | 置信度计算 (走 confidence_scorer) |
| 任务输出可结构化为 schema | 引文回链检查 (走 citation_checker) |
high severity findings 必须有 repair_actions 处理| 失败 | 回退 |
|---|---|
| 充足性门: hypothesis 不足 | 触发 Intake 重新解析, 用更细颗粒度 |
| 证据门: incident 数不足 | 继续 interview, 把 hypothesis 优先级最高的换一个角度问 |
| 命名门: 机制全部只支撑 1 张证据 | 触发 Mechanism Modeler 重新聚类, 若仍不足则降到 hypothesis |
| 一致性门: Skeptic 无法消除高严重度问题 | 强制降级为 probable; 若 confidence < 0.4, 移到 needs_more_evidence |
| 发布门: PII 未清理 | redact 后重试, 不允许跳过 |
Orchestrator 在每个阶段切换时只输出: