Double Loop Learning — AI skill for Claude Code

Double-Loop Learning | 双环思考

A cognitive base that shifts reflection from fixing errors to questioning the assumptions that produce them. Works with any LLM agent — Claude, GPT, Gemini, or custom frameworks.

让 AI agent 从"修正错误"转向"质疑产生错误的假设"。适用于任何 LLM agent — Claude、GPT、Gemini 或自定义框架。

---

What it does | 它做什么

Most AI agents operate in single-loop mode: when something fails, they adjust the behavior and try again. Double-Loop Learning adds a meta-reflection layer that examines the mental models, assumptions, and governing variables behind the behavior — and changes those when needed.

大多数 AI agent 运行在单环模式：出错时调整行为再试。双环思考增加了一个元反思层——审视行为背后的心智模型、假设和控制变量，在必要时改变这些根源。

Before (default agent) | 安装前

"Our deploys keep failing. How do we fix our CI pipeline?"

"Here are 8 steps to harden your CI pipeline: pin dependency versions, add retry logic, increase timeout buffers, add smoke tests..."

After (with Double-Loop Learning) | 安装后

"The question assumes the CI pipeline is the problem. Your deploys fail repeatedly, and each time you patch the symptom. The pattern suggests the governing assumption — 'deploy everything together on a fixed schedule' — is the source. What if you decoupled deploy units and triggered on readiness instead of schedule? That changes the system so the class of failure can't recur."

---

How it works | 工作原理

**Core cognitive shift**: From "fix the error" to "question the assumption that produced the error."

核心认知转换：从"修正错误"到"质疑产生错误的假设"。

Default mode	Target mode
Correct the behavior 修正行为	Question the governing variable that drove the behavior 质疑驱动行为的控制变量
Reflect only when things break 出问题才反思	Reflect as daily cognitive hygiene 反思是日常认知卫生
Suspend reflection when time-pressured 时间紧就跳过反思	Accelerate reflection under pressure — OODA Orientation 压力下加速反思——OODA 定向
Reflect on safe topics only 只反思安全话题	Bracket presuppositions — examine what you'd rather not 悬置预设——审视你不愿审视的
"We did a retrospective" 我们做了复