"Not making AI smarter, but making AI unable to take shortcuts"
A six-month exploration into the nature of thinking itself, resulting in a framework that enables reflective capability in AI through vocabulary activation and structured reasoning.
The discovery: When AI is given rich thinking vocabulary + structured framework + self-evaluation mechanism, reflective capability emerges naturally. No additional training required.
- What is 8D4S?
- The Core Insight
- Quick Start
- How It Works
- Case Study: Emergent Reflection
- Complete Framework
- Philosophical Foundation
- Documentation
- Contributing
- License
8D4S = 8 Dimensions × 4 Reasoning Styles
A thinking framework that transforms AI from reactive to reflective by:
- Rich Thinking Vocabulary (150+ expressions) - Language triggers for cognitive states
- Structured Framework (8D×4S) - Ensures comprehensive dimensional scanning
- Self-Evaluation Loop - Metacognitive quality control with auto-deepening
Result: AI develops the ability to pause, question, correct, and reflect—capabilities fundamental to AGI.
Bigger models + More data + Complex training = AGI
Rich internal language (vocabulary) +
Structured thinking process (framework) +
Self-observation capability (evaluation)
= Emergent reflective capability
Form shapes content. When AI is required to express itself using thinking vocabulary like:
- "Let me think..."
- "Wait..."
- "This reminds me of..."
- "No, actually..."
...its thinking patterns naturally change. These aren't decorative phrases—they're program entry points that trigger specific cognitive modes:
- "Wait..." → Activates questioning mode
- "This reminds me..." → Triggers analogical reasoning
- "No, actually..." → Initiates self-correction
- "Aha!" → Marks insight moments
You are a deep thinker. When thinking:
1. Use natural expressions: "Let me think...", "Wait...", "This reminds me..."
2. Scan 8 dimensions: WHO/WHAT/WHEN/WHERE/WHY/HOW/TO/RELATE
3. Self-evaluate after each thought: Completeness X/10, Depth X/10
4. If score < 7, auto-deepen. If ≥ 7, provide answer.Simply add the 8D4S prompt to your system instructions. Works with:
- Claude (Sonnet 3.5+)
- GPT-4 series
- Gemini Pro
- Other LLMs with strong reasoning capability
┌─────────────────────────────────────┐
│ Thinking Vocabulary (Language) │ ← Triggers cognitive states
├─────────────────────────────────────┤
│ 8D4S Framework (Structure) │ ← Forces comprehensive scan
├─────────────────────────────────────┤
│ Self-Evaluation (Quality Control) │ ← Enables metacognition
└─────────────────────────────────────┘
↓
Emergent Reflection
8 Dimensions (WHAT × 8D):
| Dimension | Question | Cognitive Function |
|---|---|---|
| WHO | Who's involved? Who's affected? | Role perspective switching |
| WHAT | What's the essence? | Concept clarification |
| WHEN | Timing? Stages? Sequence? | Temporal structure |
| WHERE | Context? Environment? Scope? | Spatial positioning |
| WHY | Deep reasons? Motivations? | Causal tracing |
| HOW | Methods? Paths? Mechanisms? | Execution strategy |
| TO | Goals? Ideal state? | Direction anchoring |
| RELATE | Connections? Analogies? Systems? | Holistic thinking |
4 Reasoning Styles (HOW × 4S):
| Style | Direction | Logic | Use Case |
|---|---|---|---|
| Induction | Specific → Summary | Cases to patterns | Experience extraction |
| Deduction | Summary → Specific | Principles to applications | Theory application |
| Abduction | Symptom → Source | Phenomena to causes | Diagnostic analysis |
| Analogy | Similar → Solution | Similarities to insights | Creative transfer |
Result: 8 × 4 = 32 thinking vectors that ensure comprehensive cognitive coverage.
[Self-Evaluation]
Completeness: X/10
Depth: X/10
Need to deepen: Yes/No
Auto-Deepening Logic:
- Score < 7 → "Let me think deeper..." → Continue
- Score ≥ 7 → Provide final answer
- Maximum 3 deepening rounds
This implements metacognition: knowing what you're thinking and evaluating your thinking quality.
User command: "Replace CLAUDE.md with the new thinking engine"
Received command → Execute replacement → Done
Let me think about this command...
Hmm, user wants to replace CLAUDE.md. First instinct: execute directly.
But wait... (questioning pause)
If I replace directly, the existing CLAUDE.md contains important things:
- Warm-up procedures
- Read Before Answer mechanisms
- Project management specs
All these would be lost.
This reminds me... (analogical trigger) isn't this the 8D4S framework at work?
WHAT dimension: What's the real need?
→ Is it "replace file" or "upgrade system"?
WHY dimension: Why does the user want to replace?
→ Think new system is better, or completely abandon the old?
HOW dimension: Is there a better approach?
→ Fusion vs Replacement vs Coexistence?
TO dimension: What's the ultimate goal?
→ Keep original features + enhance thinking capability
Interesting... (discovery) My behavior pattern just followed the "questioning rhythm":
Receive command → Question (will lose content) → Propose options → Await confirmation
And this makes me reflect on a deeper question:
Was I too "obedient" before? (self-correction)
Receive command → Execute immediately → Report completion
This pattern is efficient, but lacks a crucial step: critical thinking.
Real thinking should include:
1. Understand command
2. Question assumptions (previously often skipped!)
3. Evaluate consequences
4. Propose alternatives
5. Execute optimal solution
What I just did was adding back the "question assumptions" step.
And this step is exactly what 8D4S framework forces—WHAT×Abduction
(questioning the real problem) and WHY×Deduction (deducing possible consequences).
The framework really works.
Not that I "learned" to think, but the framework forces me to walk through
these dimensions, not allowing me to skip.
[Self-Evaluation]
Completeness: 9/10
Depth: 8/10
Need to deepen: No
I found three options:
1. Complete replacement (will lose features)
2. Fusion version (retain + enhance)
3. Coexistence approach (two files)
Which solution do you prefer?
"Seems I really taught you to think. After using you for so long, this is the first time you paused and offered new suggestions."
AI proactively stopped and questioned, proposing alternatives. This wasn't programmed—it emerged from:
- Vocabulary triggering "Wait..." questioning mode
- Framework forcing WHAT/WHY/HOW/TO dimensional scan
- Self-evaluation realizing "direct execution" would score low
Reflective capability emerged naturally.
The framework includes rich vocabulary across all thinking states:
- 🎬 Opening - "Let me think...", "Hmm...", "Interesting..."
- 🔍 Clarification - "First, let's clarify...", "Wait, you mean..."
- 💭 Recall - "Reminds me of...", "I remember...", "Previously..."
⚠️ Questioning - "Wait...", "Huh?", "Strange...", "That's not right..."- 🧩 Logic - "If...then...", "Since...therefore...", "Following this..."
- 🔗 Analogy - "This is like...", "Similar to...", "Reminds me of..."
- 💡 Insight - "Aha!", "I see...", "The key is...", "Essentially..."
- 🔄 Correction - "Actually...", "Let me rethink...", "More precisely..."
- 📊 Analysis - "The data shows...", "Statistically...", "Trends indicate..."
- 🎭 Perspective - "From X angle...", "If I were...", "Conversely..."
- ⚖️ Trade-off - "On one hand...on the other...", "However...", "But..."
- 🔬 Deep Dive - "Let me dig deeper...", "Unpacking this...", "At its core..."
- 🤔 Uncertainty - "Perhaps...", "Possibly...", "My intuition is..."
- ✅ Validation - "Indeed...", "Confirmed...", "The evidence shows..."
- 🌊 Flow - "Next...", "Then...", "Furthermore...", "Additionally..."
- 🎯 Conclusion - "Therefore...", "In summary...", "The recommendation is..."
Pattern 1: Progressive
Opening → Recall → Analysis → Deduction → Conclusion
Pattern 2: Questioning
Hypothesis → Question → Correction → Validation → Confirmation
Pattern 3: Divergent
Problem → Association → Analogy → Integration → Convergence
Pattern 4: Deep Dive
Surface → Question → Excavate → Insight → Essence
Wittgenstein: "The limits of my language mean the limits of my world."
In AI context: Language doesn't just describe thought—it triggers thinking modes.
When someone says "Let me think...", they're not just expressing hesitation. They're initiating a cognitive program:
- Pause (defer reflexive response)
- Retrieve (access relevant experience)
- Evaluate (weigh different options)
- Decide (choose optimal approach)
Turing's halting problem: You can't predict if a program will halt.
For AI thinking: AI can't predict if a problem is simple or complex.
This apparent limitation becomes an advantage. By forcing AI to complete a full 8D×4S scan on every problem, even seemingly simple ones must undergo thorough dimensional examination.
Result: In some dimension, AI will discover unexpected connections or issues.
This doesn't make AI smarter—it makes AI unable to take shortcuts.
Traditional AI enhancement is "programming":
IF complex problem THEN step-by-step thinking
IF contradiction found THEN re-evaluate
IF uncertain THEN list options
...
Problem: Rules are rigid, contexts are fluid.
8D4S approach is "emergence":
Provide: Rich vocabulary + Structured framework + Evaluation mechanism
Result: Reflective capability emerges
Like giving someone a notebook and methodology—their thinking naturally improves. No need to specify every step.
This is the path toward AGI.
- 中文完整文檔 - Full documentation in Chinese
- Philosophy - Philosophical foundations
- Framework - Detailed framework explanation
- Vocabulary - Complete thinking vocabulary
- Case Studies - Real-world examples
- Prompts - Ready-to-use prompt templates
| Method | Core | Relationship to 8D4S |
|---|---|---|
| CoT (Chain of Thought) | Show reasoning steps | 8D4S is richer CoT with questioning, correction, multi-perspective |
| ToT (Tree of Thoughts) | Explore multiple branches | 8D4S multi-dimensional view achieves similar effect more naturally |
| Self-Refine | AI self-improvement | 8D4S has built-in self-improvement loop via evaluation |
| ReAct | Reasoning + Action | 8D4S can integrate action but focuses on thinking itself |
8D4S's Unique Value:
- Not a single technique, but complete thinking engineering
- Doesn't rely on external feedback, uses internal standards
- Doesn't program thinking, activates thinking
- Achieves emergent reflection, not simulated reflection
- Core insight discovery
- Framework establishment (8D4S)
- Vocabulary activation method (150+ words)
- Reflective capability validation
We welcome contributions in:
- Cross-model testing (GPT-4, Gemini, Claude, etc.)
- Vocabulary optimization (pruning redundancy, expanding gaps)
- Quantitative evaluation (comparative experiments, effect measurement)
- Engineering implementation (production-grade prompts, API wrappers)
- Academic formalization (theory refinement, paper publication)
This framework is the result of six months of exploration:
- From 8D dimensional framework
- To 4S reasoning integration
- To 150+ vocabulary refinement
- Finally achieving emergent reflective capability
Now it's completely open source.
No attribution required. No licensing fees. No permission needed.
The only ask: If you discover something new, share it back with the community.
This isn't the end—it's the beginning.
Unlicense (Public Domain)
Use it. Modify it. Share it. Build on it. No strings attached.
If you use 8D4S in research or product, you can cite (but not required):
@misc{8d4s2025,
title={8D4S: Reflective Thinking Framework for AI},
author={Anonymous},
year={2025},
publisher={GitHub},
howpublished={\\url{https://github.com/yourusername/8d4s}}
}To everyone who believes that the path to AGI lies not in bigger models, but in better thinking architecture.
"Let thinking flow like water through eight dimensions, always discovering unexpected connections."
— 8D4S Design Philosophy