complex-reasoning
About
This skill enables structured problem-solving using frameworks like Chain-of-Thought and Tree-of-Thought for multi-step analysis. It's designed for complex tasks such as debugging, root cause analysis, and architectural decisions. The skill leverages extended thinking and tools like Bash, Grep, and Edit to systematically break down and solve intricate problems.
Quick Install
Claude Code
Recommended/plugin add https://github.com/majiayu000/claude-skill-registrygit clone https://github.com/majiayu000/claude-skill-registry.git ~/.claude/skills/complex-reasoningCopy and paste this command in Claude Code to install this skill
Documentation
Complex Reasoning Skill
Structured reasoning frameworks for systematic problem solving, leveraging extended thinking capabilities for deep analysis.
When to Use
- Debugging complex issues with multiple potential causes
- Architecture decisions requiring trade-off analysis
- Root cause analysis for production incidents
- Performance optimization with multiple variables
- Security vulnerability assessment
- Code refactoring with many dependencies
Reasoning Frameworks
Chain-of-Thought (CoT)
Linear step-by-step reasoning for sequential problems.
## Chain-of-Thought Analysis
**Problem**: [State the problem clearly]
**Step 1: Understand the Context**
- What do we know?
- What are the constraints?
- What is the expected outcome?
**Step 2: Identify Key Components**
- Component A: [description]
- Component B: [description]
- Interactions: [how they relate]
**Step 3: Analyze Each Component**
- Component A analysis...
- Component B analysis...
**Step 4: Synthesize Findings**
- Key insight 1
- Key insight 2
**Step 5: Formulate Solution**
- Recommended approach
- Rationale
- Trade-offs
**Conclusion**: [Final recommendation with confidence level]
Tree-of-Thought (ToT)
Branching exploration for problems with multiple solution paths.
## Tree-of-Thought Exploration
**Root Problem**: [Problem statement]
### Branch 1: Approach A
├── Pros: [List advantages]
├── Cons: [List disadvantages]
├── Feasibility: [High/Medium/Low]
├── Sub-branch 1.1: [Variation]
│ └── Outcome: [Expected result]
└── Sub-branch 1.2: [Variation]
└── Outcome: [Expected result]
### Branch 2: Approach B
├── Pros: [List advantages]
├── Cons: [List disadvantages]
├── Feasibility: [High/Medium/Low]
└── Sub-branches: [...]
### Branch 3: Approach C
├── Pros: [...]
├── Cons: [...]
└── Feasibility: [...]
### Evaluation Matrix
| Approach | Feasibility | Impact | Risk | Score |
|----------|-------------|--------|------|-------|
| A | High | Medium | Low | 8/10 |
| B | Medium | High | Med | 7/10 |
| C | Low | High | High | 5/10 |
**Selected Path**: Branch [X] because [reasoning]
MECE Framework
Mutually Exclusive, Collectively Exhaustive analysis.
## MECE Analysis
**Problem Space**: [Define the complete problem]
### Category 1: [Mutually exclusive category]
- Sub-element 1.1
- Sub-element 1.2
- Sub-element 1.3
### Category 2: [Mutually exclusive category]
- Sub-element 2.1
- Sub-element 2.2
### Category 3: [Mutually exclusive category]
- Sub-element 3.1
- Sub-element 3.2
- Sub-element 3.3
**Completeness Check**:
- [ ] Categories are mutually exclusive (no overlap)
- [ ] Categories are collectively exhaustive (cover all cases)
- [ ] Each sub-element belongs to exactly one category
**Priority Matrix**:
| Category | Urgency | Impact | Action |
|----------|---------|--------|--------|
| 1 | High | High | Now |
| 2 | Medium | High | Next |
| 3 | Low | Medium | Later |
Hypothesis-Driven Debugging
Systematic approach to debugging complex issues.
## Hypothesis-Driven Debug Session
**Symptom**: [Observed behavior]
**Expected**: [What should happen]
**Environment**: [Relevant context]
### Hypothesis 1: [Most likely cause]
**Evidence For**:
- [Supporting observation 1]
- [Supporting observation 2]
**Evidence Against**:
- [Contradicting observation]
**Test**: [How to validate]
**Result**: [Confirmed/Refuted]
### Hypothesis 2: [Second most likely]
**Evidence For**:
- [...]
**Evidence Against**:
- [...]
**Test**: [...]
**Result**: [...]
### Root Cause Identified
**Cause**: [Confirmed root cause]
**Evidence Chain**: [How we proved it]
**Fix**: [Remediation steps]
**Prevention**: [How to prevent recurrence]
Code Analysis Patterns
Dependency Analysis
## Dependency Analysis: [Component Name]
### Direct Dependencies
| Dependency | Version | Purpose | Risk Level |
|------------|---------|---------|------------|
| dep-a | 2.3.1 | Auth | Low |
| dep-b | 1.0.0 | Data | Medium |
### Transitive Dependencies
- Total: [N] packages
- Security vulnerabilities: [N]
- Outdated: [N]
### Dependency Graph
[component] ├── dep-a │ ├── sub-dep-1 │ └── sub-dep-2 └── dep-b └── sub-dep-3
### Risk Assessment
1. **High Risk**: [Dependencies with known issues]
2. **Medium Risk**: [Outdated or unmaintained]
3. **Low Risk**: [Stable, well-maintained]
### Recommendations
1. [Action item 1]
2. [Action item 2]
Impact Analysis
## Impact Analysis: [Proposed Change]
### Affected Components
| Component | Impact Type | Severity | Test Required |
|-----------|-------------|----------|---------------|
| Service A | Direct | High | Yes |
| Service B | Indirect | Medium | Yes |
| Client C | Downstream | Low | Optional |
### Risk Assessment
- **Breaking Changes**: [List any]
- **Performance Impact**: [Expected effect]
- **Data Migration**: [Required/Not required]
### Rollback Plan
1. [Step 1]
2. [Step 2]
3. [Verification]
### Recommendation
[Go/No-Go with reasoning]
Integration with Extended Thinking
When using these frameworks with extended thinking:
# Enable extended thinking for complex reasoning
response = client.messages.create(
model="claude-opus-4-5-20250514",
max_tokens=16000,
thinking={
"type": "enabled",
"budget_tokens": 15000 # Higher budget for complex reasoning
},
system="""You are a systematic problem solver. Use structured
reasoning frameworks like Chain-of-Thought, Tree-of-Thought,
or MECE analysis as appropriate for the problem.""",
messages=[{
"role": "user",
"content": "Analyze this architecture decision using ToT..."
}]
)
Best Practices
- Choose the right framework: CoT for linear problems, ToT for branching decisions
- Document your reasoning: Makes it reviewable and repeatable
- Validate assumptions: Each step should build on verified facts
- Consider alternatives: Always explore at least 2-3 approaches
- Quantify when possible: Use metrics to compare options
- Time-box exploration: Set limits on analysis depth
See Also
- [[extended-thinking]] - Enable deep reasoning capabilities
- [[deep-analysis]] - Analytical templates
- [[debugging]] - General debugging patterns
- [[testing]] - Validation strategies
GitHub Repository
Related Skills
sglang
MetaSGLang is a high-performance LLM serving framework that specializes in fast, structured generation for JSON, regex, and agentic workflows using its RadixAttention prefix caching. It delivers significantly faster inference, especially for tasks with repeated prefixes, making it ideal for complex, structured outputs and multi-turn conversations. Choose SGLang over alternatives like vLLM when you need constrained decoding or are building applications with extensive prefix sharing.
evaluating-llms-harness
TestingThis Claude Skill runs the lm-evaluation-harness to benchmark LLMs across 60+ standardized academic tasks like MMLU and GSM8K. It's designed for developers to compare model quality, track training progress, or report academic results. The tool supports various backends including HuggingFace and vLLM models.
langchain
MetaLangChain is a framework for building LLM applications using agents, chains, and RAG pipelines. It supports multiple LLM providers, offers 500+ integrations, and includes features like tool calling and memory management. Use it for rapid prototyping and deploying production systems like chatbots, autonomous agents, and question-answering services.
llamaguard
OtherLlamaGuard is Meta's 7-8B parameter model for moderating LLM inputs and outputs across six safety categories like violence and hate speech. It offers 94-95% accuracy and can be deployed using vLLM, Hugging Face, or Amazon SageMaker. Use this skill to easily integrate content filtering and safety guardrails into your AI applications.
