跳到正文
TakoAPI
返回技能列表

chaos-pivot

由 manecharo 提供

AI & LLMs
418下载量由 clawskills.sh 提供
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

Prevents LLMs from sunk-cost pushing broken solutions.

描述

该技能的功能

Addresses a silent failure mode in agentic AI where an LLM keeps pushing a broken approach past the point of reason. When an agent is stuck, looping, or failing repeatedly, it forces a structured stop: name the failed assumption, generate three genuinely different alternatives using inversion, analogy, and brute-force simplicity, then probe each for signal. Loops through a design-thinking cycle until a viable path is found or the problem is escalated with a clear failure diagnosis.

Most agents default to cosmetic retries or parameter tweaks around a broken core assumption; this skill forces genuine strategic divergence by requiring each alternative to violate at least one assumption of the previous approach.

何时使用

  • Agent retrying the same broken API call three times in a row
  • Debugging session where every attempted fix hits the same compiler error
  • Code generation task looping on the same logic structure without progress
  • Automation workflow timing out at the same step across multiple runs
  • Data extraction script returning wrong results despite incremental parameter tweaks

示例工作流

输入

User asks: fix this script that keeps failing to parse the API response

智能体
  1. 1

    Recognize the trigger: same approach attempted 3 times with no material improvement

  2. 2

    Write a dead end declaration naming the failed assumption and what it revealed

  3. 3

    Generate 3 alternatives: Inversion (parse locally instead of calling the API), Analogy (treat the response like a stream, not a document, borrowing from log-parsing patterns), Reduction (hardcode a known-good response shape and verify just the extraction logic)

  4. 4

    Probe each alternative for signal without needing to fully complete each one

  5. 5

    Run a third-party audit on the strongest result, checking for scope drift and edge cases before declaring done

输出

Working parser using the local-computation approach, with audit confirming it handles the actual response structure

安装

clawhub install manecharo/chaos-pivot

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