中文 | English
Upgrade AI from answer-generator to cognitive operating system
metacog is a single Claude Code skill that packs 8 metacognitive capabilities into one /metacog command. Instead of just answering your questions, it helps you think better about thinking itself.
Most AI interactions follow: User asks → AI answers. The real bottleneck is rarely "what's the answer" — it's "am I even asking the right question?" metacog flips the script: it analyzes the thinking process itself.
# Personal (all your projects)
mkdir -p ~/.claude/skills/metacog
cp SKILL.md ~/.claude/skills/metacog/
# Or project-level (this project only)
mkdir -p .claude/skills/metacog
cp SKILL.md .claude/skills/metacog/Restart Claude Code after installation.
| Keyword | Skill | What it does | Theory |
|---|---|---|---|
cognition |
Metacognition | Diagnoses thinking errors, biases, logic gaps | Flavell, Dunning-Kruger |
learning |
Meta-Learning | Designs learning paths, finds the critical 20% | Feynman, Pareto, Ericsson |
decision |
Meta-Decision | Analyzes trade-offs, runs pre-mortems | Kahneman, prospect theory |
structure |
Meta-Structure | Builds MECE logical skeletons | Minto Pyramid, McKinsey |
feedback |
Meta-Feedback | Traces behavior→mechanism→outcome chains | Nelson & Narens |
review |
Meta-Review | Extracts SOPs from past events via 5 Whys | Reflexion, Argyris |
emotion |
Meta-Emotion | Separates emotional signal from noise | Gottman, Gross, Slovic |
communication |
Meta-Communication | Decodes meta-messages, fixes intent-impact gaps | Bateson, Watzlawick, NVC |
/metacog I study 8 hours a day but my exam scores are terrible
→ Routing → learning
→ (identifies fluency illusion, designs active recall path)
/metacog decision Should I quit my job to start a business?
→ (decision principles, trade-off matrix, pre-mortem, fallback plan)
/metacog all I keep procrastinating and feel terrible about it
→ Runs emotion + cognition + feedback → Synthesizes insights
Every skill output always includes these 5 guarantees:
Every skill is grounded in peer-reviewed academic frameworks, not vibes.
Nelson & Narens Monitor-Control (1990) — The operating backbone
Two cognitive levels interact:
- Monitor: meta-level observes the object-level ("How am I doing?")
- Control: meta-level modifies the object-level ("Switch strategy")
Every metacog skill follows this: first monitor the user's thinking, then provide control actions.
Kahneman Dual Process (2011) — Why we need metacognition
- System 1: Fast, automatic, intuitive, error-prone
- System 2: Slow, deliberate, rational, energy-consuming
Metacognition = using System 2 to audit System 1.
Gottman Meta-Emotion (1996) — Emotions about emotions
- Emotion-Coaching: recognize → name → understand
- Emotion-Dismissing: deny → suppress → avoid
The emotion skill implements the coaching approach: emotions as data, not noise.
Argyris Double-Loop Learning (1978) — Question the model
- Single-loop: action failed → fix the action
- Double-loop: action failed → question the assumption → fix the mental model
The review skill asks not just "what went wrong" but "what belief was wrong."
| Decision | Why |
|---|---|
| Single file | Users remember one command: /metacog. Router handles complexity. |
| Bilingual headers | English for LLM precision; Chinese for intuitive understanding. |
| Confidence 1-10 | Combats Dunning-Kruger for both AI and user. |
| Anti-flattery | Every output must survive: "Delete all compliments — what's left?" |
| Reusable principles | Each interaction extracts one transferable insight. |
| cognition | emotion | |
|---|---|---|
| Target | Thinking errors | Emotional distortion |
| Core Q | "Is my reasoning valid?" | "What am I feeling? How does it bend judgment?" |
| Anti-pattern | "You're right" (flattery) | "Don't overthink it" (denial) |
| structure | communication | |
|---|---|---|
| Target | Information architecture | Interpersonal dynamics |
| Core Q | "Is this organized clearly?" | "What is really being said?" |
| Anti-pattern | "Everything matters" (flat) | "Just be honest" (naive) |
These outputs are explicitly forbidden:
| Pattern | Bad | Good |
|---|---|---|
| Flattery | "Great question!" | Jump to analysis |
| Chicken soup | "Believe in yourself!" | Concrete steps |
| Hedging | "It depends" | List variables + conditionals |
| Emotion denial | "Don't overthink it" | Name the emotion, trace effects |
| Naive advice | "Just be honest" | Analyze dynamics, give phrasing |
| Flattening | "All equally important" | Priority ranking |
| Hallucination | Fabricating answers | State what's missing, ask |
Core: Flavell (1979), Schraw & Dennison MAI (1994), Nelson & Narens (1990), Kahneman (2011), Kruger & Dunning (1999), Ericsson (1993), Gottman (1996), Gross (1998), Slovic (2007), Bateson (1972), Watzlawick (1967), Schulz von Thun (1981), Rosenberg (2003), Argyris (1978), Minto (2009), Shinn (2023).
Projects: atscub/metacognition · thinking-partner · LangGPT
Open docs/index.html in your browser for the full interactive documentation with visual diagrams.
Think about your thinking. Learn about your learning. Decide about your deciding.
MIT License
Maintained by TalentOS · 需要定制 Agent / 技能开发？找我们定制 / Work with us · talentos@agentbay.asia