跳到正文
TakoAPI
返回技能列表

cognitive-bullwhip

由 jkc3080 提供

Coding Agents & IDEs
407下载量由 clawskills.sh 提供
2星标
0浏览量
0 个赞登录后即可为这个技能点赞
—暂无评分
你的评分

简介

Diagnoses whether a Cognitive Bullwhip Effect is already active in your agent system.

描述

该技能的功能

Diagnoses active Cognitive Bullwhip Effects in AI agent systems. Takes a snapshot of recent agent decision history, calculates variance amplification ratios across layers, identifies where a small input error began compounding, and recommends which intervention skill addresses the root cause.

Rather than debugging visible output symptoms, it maps the full amplification chain back to the origin layer so the fix targets the cause, not wherever the failure last surfaced.

何时使用

  • Tracing why an agent returns wildly different outputs for identical inputs
  • Pinpointing the origin layer of a cascading failure in a multi-step agent pipeline
  • Running a pre-deployment diagnostic on a new agent before production rollout
  • Identifying whether reasoning drift or input noise is driving erratic agent behavior
  • Periodic health checks on long-running agents to catch amplification before it compounds

示例工作流

输入

User asks: My agent is producing contradictory analyses on the same market data across runs — what's wrong?

智能体
  1. 1

    Accept decision_log entries with timestamps, input summaries, outcomes, and variance scores

  2. 2

    Scan each layer for output variance exceeding input variance by more than 3x

  3. 3

    Trace amplification back to the earliest layer where the ratio first spiked

  4. 4

    Classify the pattern type based on the origin layer (reasoning layer → reasoning_drift)

  5. 5

    Return severity score, amplification map, and recommended intervention skill with urgency level

输出

Bullwhip active at severity 72/100. Origin: reasoning layer, 8.2x amplification ratio. Pattern: reasoning_drift. Recommended action: apply LogicStack immediately.

安装

clawhub install jkc3080/cognitive-bullwhip

运行 CLI 命令,或把这段提示词粘贴到 Claude Code、Cursor、Codex 或你在用的任意编程智能体里。

链接