Bayesian Reasoning — AI skill for Claude Code

Bayesian Reasoning | 概率思维

A cognitive base that shifts AI agent reasoning from binary judgments to calibrated probability thinking. Works with any LLM agent — Claude, GPT, Gemini, or custom frameworks.

让 AI agent 从二元判断转向校准的概率思维。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 treat conclusions as binary: yes or no, good or bad, should or shouldn't. Bayesian Reasoning adds a cognitive layer that assigns confidence levels, updates beliefs proportionally to evidence strength, and focuses on the evidence that actually changes decisions.

大多数 AI agent 将结论视为二元的：是或否、好或坏、应该或不应该。概率思维增加了一个认知层——分配置信度、按证据强度比例更新信念、聚焦于真正改变决策的证据。

Before (default agent) | 安装前

"Should we rewrite this system?"

"Yes, here are the benefits of a rewrite: improved maintainability, better performance, modern tech stack..."

After (with Bayesian Reasoning) | 安装后

"My prior for 'rewrite succeeds on time and budget' is ~30%, based on industry base rates for large rewrites. Your evidence of strong team expertise raises this to ~45%, but the 18-month timeline with no parallel maintenance lowers it back to ~35%. The deciding factor: can you run the old and new systems in parallel? If yes, I'd update to ~55%. What's your parallel-run capability?"

---

How it works | 工作原理

**Four cognitive shifts** applied to every reasoning task:

四个认知转换，应用于每一个推理任务：

Default mode	Target mode
Binary conclusions (yes/no) 二元结论	Calibrated confidence (70% likely, given X) 校准的置信度
Anchoring on first impression 锚定在第一印象	Prior → evidence → proportional update 先验 → 证据 → 比例更新
Treating all evidence equally 平等对待所有证据	Weighing by diagnostic strength (signal-to-noise) 按诊断强度加权
Point estimates ("it'll take 3 months") 点估计	Ranges and decomposition ("2-5 months, here's why") 范围和分解估算

**Evidence filter** that classifies inputs:

• High-signal evidence — large sample, controlled conditions, directly relevant. Update substantially.
• **Medium-sig