# Install in 10 seconds:
curl -sSL https://raw.githubusercontent.com/TheNight-Watch/deepthink-skill/main/install.sh | bashStop letting AI think for you. Start thinking deeper with AI.
Install Β· How It Works Β· Commands Β· Setup Guide
- Three-stage workflow: Understand β Research β Socratic Thinking
- Three independent sources: X/Twitter (xurl) + Perplexity AI (2-round) + 11 curated RSS feeds
- Belief tracking: Records how your views evolve across thinking sessions
- Built-in setup wizard:
/deepthink setupguides API key configuration - Cross-session memory: Picks up where you left off, no context lost
AI makes it easy to get answers. Too easy. You ask, it answers, you move on.
Over time, you stop forming your own opinions. You outsource judgment. You lose the muscle of deep thinking.
DeepThink flips the script: instead of giving you answers, it gives you better questions.
You pick a topic β DeepThink researches it β Then coaches your thinking
| Stage | What Happens | Your Role |
|---|---|---|
| 1. Understand | DeepThink asks why you care about this topic | Share your motivation |
| 2. Research | Pulls insights from X/Twitter, Perplexity AI, and 11 curated RSS feeds | Sit back (async) |
| 3. Think | Socratic questioning based on research findings | Think, argue, evolve |
The magic is in Stage 3: DeepThink doesn't summarize β it challenges. It asks follow-up questions, presents counter-arguments, tracks how your beliefs evolve over time.
curl -sSL https://raw.githubusercontent.com/TheNight-Watch/deepthink-skill/main/install.sh | bashPaste this into Claude Code:
Install the DeepThink skill from https://github.com/TheNight-Watch/deepthink-skill β copy the
skill/folder to~/.claude/skills/deepthink/and thecommand/deepthink.mdto~/.claude/commands/deepthink.md, then create~/.claude/deepthink-data/with subdirectoriestopics/,research/,sessions/and anindex.jsoncontaining{"topics":[]}.
# Clone
git clone https://github.com/TheNight-Watch/deepthink-skill.git
cd deepthink-skill
# Copy skill files
cp -r skill/ ~/.claude/skills/deepthink/
cp command/deepthink.md ~/.claude/commands/
# Create data directory
mkdir -p ~/.claude/deepthink-data/{topics,research,sessions}
echo '{"topics":[]}' > ~/.claude/deepthink-data/index.jsonThen run /deepthink setup in Claude Code to configure API keys.
After installing, run /deepthink setup in Claude Code. It will guide you through:
DeepThink uses Perplexity's sonar model for deep research.
- Go to perplexity.ai/settings/api
- Create an API key
- Paste it when
/deepthink setupasks
Stored locally at
~/.claude/deepthink-data/.envβ never leaves your machine.
xurl lets DeepThink search X/Twitter for real-time expert opinions.
# macOS
brew install xurl
# Then authenticate
xurl auth login --app deepthinkWithout xurl, DeepThink still works β it uses Perplexity + RSS only.
DeepThink comes with 11 curated AI-focused RSS sources including:
- Simon Willison's Weblog, Lil'Log (Lilian Weng), Import AI
- Latent Space, One Useful Thing (Ethan Mollick), Every
- Ben's Bites, AINews, Interconnects, Ahead of AI, The Batch
Want to add your own? Edit ~/.claude/skills/deepthink/sources.md.
| Command | What it does |
|---|---|
/deepthink setup |
First-time configuration wizard |
/deepthink new: {topic} |
Start thinking about a new topic |
/deepthink list |
See all your active thinking topics |
/deepthink research {topic} |
Run 3-source deep research |
/deepthink read {topic} |
View the latest research article |
/deepthink chat {topic} |
Start a Socratic thinking session |
/deepthink continue {topic} |
Continue from where you left off |
/deepthink summary {topic} |
See your full thinking timeline |
/deepthink archive {topic} |
Archive a completed topic |
DeepThink doesn't just Google things. It pulls from three independent source types:
X/Twitter (xurl) β Real-time expert opinions, debates
Perplexity AI (2 rounds) β Deep research with citations
RSS (11 curated feeds) β Long-form analysis from top AI writers
Then synthesizes everything into a research article organized by arguments (not by source), with cross-source validation, contradictions highlighted, and thinking questions derived from gaps in the evidence.
DeepThink acts as a thinking coach, not a lecturer:
- π Explores your current position
- π¬ Deep-dives into vague statements ("what do you mean by 'autonomy'?")
- β‘ Challenges with counter-evidence from research
- π Connects scattered ideas into coherent arguments
- π Summarizes your thinking evolution over time
It tracks your belief changes across sessions β so you can see how your thinking evolved.
Session 1: "AI agents will replace most jobs"
Session 3: "AI agents will transform jobs, but human judgment becomes more valuable"
Session 5: "The real question isn't replacement vs augmentation β it's who controls the agents"
~/.claude/
βββ skills/deepthink/
β βββ SKILL.md # Main skill definition
β βββ sources.md # 11 curated RSS + 9 X/Twitter KOLs
β βββ references/
β βββ stage1_understanding.md # Topic onboarding flow
β βββ stage2_research.md # Research execution manual
β βββ stage3_thinking.md # Socratic coaching framework
βββ commands/
β βββ deepthink.md # Slash command router
βββ deepthink-data/
βββ .env # API keys (local only)
βββ index.json # Topic index
βββ topics/ # Topic records
βββ research/ # Research articles
βββ sessions/ # Thinking session logs
| Feature | DeepThink | Regular AI Chat |
|---|---|---|
| Research depth | 3 independent sources, 2-round Perplexity | Single model response |
| Thinking guidance | Socratic questioning | Gives you the answer |
| Belief tracking | Records how your views evolve | No memory |
| Source quality | 11 curated expert RSS feeds | Generic web search |
| Cross-validation | Highlights contradictions between sources | No cross-checking |
- Three-stage workflow (Understand β Research β Think)
- X/Twitter + Perplexity + RSS integration
- Socratic questioning with belief tracking
- Cross-session thinking timeline
- Scheduled research (auto-research on interval)
- Multi-language support
- Custom source libraries (beyond AI)
- Export thinking journal to Markdown/Notion
PRs welcome! Areas where help is needed:
- New source libraries β extend beyond AI topics (finance, science, politics)
- Language support β the skill currently works best in Chinese and English
- Thinking frameworks β alternative coaching styles beyond Socratic method
MIT β see LICENSE
Built for humans who refuse to stop thinking.
Star β this repo if you believe AI should make us think more, not less.