deepthink

Smart deep thinking skill for Claude Code - automatically adapts depth based on question complexity.

What is this?

This skill triggers intelligent, adaptive reasoning. It automatically detects question complexity and chooses the appropriate response depth - quick answers for simple questions, deep analysis for complex ones.

Key Features

• Automatic Complexity Detection - Smart mode selection
• Quick Mode - Direct answers for simple questions
• Deep Mode - Full analysis for complex problems
• Adaptive Output - Not every question needs full protocol
• Selective Self-Questioning - Only relevant questions

Installation

Method 1: Clone to skills directory

git clone https://github.com/wxhou/deepthink.git ~/.claude/skills/deepthink

Method 2: Using Plugin

/plugin install deepthink-marketplace

Usage

/deepthink Your question here

How It Works

Step 1: Complexity Assessment

First, assess whether the question needs deep analysis:

Quick Mode (skip full protocol):

• Simple facts ("What day is today?")
• Confirmations ("Does this file exist?")
• Single tasks ("Run npm install")
• Yes/No questions

Deep Mode (follow protocol):

• Multiple factors to consider
• Trade-offs to evaluate
• No clear "right answer"
• Requires research
• Decision-making

Step 2: Adaptive Response

Question Type	Response Style
Simple fact	Direct answer
How-to guide	Steps + key points
Decision	Pros/cons + recommendation
Analysis	Core insight + evidence

Output Examples

Quick Mode

/deepthink 今天周几？
→ 今天是周五。

Deep Mode

/deepthink 应该选择 PostgreSQL 还是 MongoDB？

## 分析

### 核心问题
数据库选型需要根据场景评估

### 关键分析
- PostgreSQL: 强一致性、复杂查询、JSON支持
- MongoDB: 灵活schema、高写入、文档存储

### 结论
根据场景选择：
- 金融/交易 → PostgreSQL
- 内容/日志 → MongoDB

### 置信度: Medium

FAQ

Q: Does every question need full analysis?

A: No! The skill automatically detects complexity. Simple questions get quick answers.

Q: What if I'm not sure?

A: When uncertain, choose Deep Mode to be safe.

Q: Do I need sequentialthinking?

A: It's built into Claude Code - no installation needed.

---

Testing

Run evals to compare skill versions:

# 1. Snapshot current skill as baseline
cp -r deepthink ~/.claude/skills/deepthink-workspace/skill-snapshot/

# 2. Edit skill, then run evals (use skill-creator workflow)
# Evals are in evals/evals.json — 12 test cases covering quick/medium/high complexity

# 3. Compare outputs: new should beat baseline on quick mode speed

File Structure

deepthink/
├── SKILL.md           # Entry point with auto language detection
├── references/
│   ├── zh.md          # Chinese version (中文版)
│   └── en.md          # English version
├── evals/
│   └── evals.json     # 12 test cases
└── scripts/
    └── run_evals.py   # Test runner

Language Detection: Automatically detects from user's prompt — Chinese characters → references/zh.md, otherwise → references/en.md.

Version

Version: 5.0.0 Author: wxhou