context-engineering-kit（上下文工程套件）

Context Engineering Kit

A collection of advanced context engineering techniques and patterns designed to improve AI agent results while minimizing token consumption.

When to Use This Skill

• Improving AI output quality systematically
• Reducing token usage in complex tasks
• Implementing structured reasoning patterns
• Multi-agent code review workflows
• Spec-driven development processes
• When standard prompting isn't enough

Core Techniques

1. Reflexion Pattern

Feedback loops that improve output by 8-21% across tasks.

How it works:

1. Generate initial response
2. Self-evaluate against criteria
3. Identify improvement areas
4. Generate refined response
5. Repeat until quality threshold met

Use when:

• Writing complex code
• Creating documentation
• Solving multi-step problems

2. Spec-Driven Development

Based on GitHub Spec Kit and OpenSpec frameworks.

Process:

# Specification Document

## Requirements
[Clear, testable requirements]

## Acceptance Criteria
[Specific success conditions]

## Constraints
[Limitations and boundaries]

## Examples
[Input/output pairs]

Benefits:

• Clearer requirements
• Testable outputs
• Reduced ambiguity

3. Subagent-Driven Development

Competitive generation with quality gates.

Architecture:

┌─────────────────────────┐
│   Orchestrator Agent    │
├─────────────────────────┤
│  ┌─────┐  ┌─────┐      │
│  │Gen 1│  │Gen 2│ ...  │
│  └─────┘  └─────┘      │
├─────────────────────────┤
│    Quality Gate Agent   │
└─────────────────────────┘

Flow:

1. Multiple agents generate solutions
2. Quality agent evaluates each
3. Best solution selected/merged
4. Iterative refinement

4. First Principles Framework (FPF)

Hypothesis-driven decision making.

Structure:

Observation → Hypothesis → Test → Conclusion

Application:

• Debugging complex issues
• Architecture decisions
• Technology selection

5. Kaizen (Continuous Improvement)

Systematic iterative enhancement.

Cycle:

Plan → Do → Check → Act → Repeat

Plugin Categories

Reasoning Enhancement

• Reflexion: Self-evaluation loops
• Chain of Thought: Step-by-step reasoning
• Tree of Thoughts: Branching exploration

Code Quality

• Multi-Agent Review: Parallel code analysis
• Security Audit: Vulnerability detection
• Performance Analysis: Optimization suggestions

Development Process

• Spec-Driven: Requirements-first approach
• TDD Support: Test-first workflows
• Documentation: Auto-generated docs

Meta-Skills

• Plugin Development: Create new plugins
• Workflow Composition: Combine techniques
• Performance Tuning: Optimize patterns

How to Use

Basic: Reflexion Loop

Review my code with reflexion:

[paste code]

Requirements:
- Error handling
- Performance
- Readability

Spec-Driven Task

Create a spec for: User authentication system

Then implement following the spec.

Multi-Agent Review

Review this PR with multiple perspectives:
- Security focus
- Performance focus
- Maintainability focus

[paste code or PR link]

Token Efficiency Tips

1. Structured Prompts

## Context
[Brief, relevant context only]

## Task
[Clear, specific task]

## Output Format
[Expected structure]

2. Progressive Disclosure

• Start with essential info
• Add details only when needed
• Remove redundant context

3. Pattern Libraries

• Reuse proven patterns
• Reference by name
• Avoid repeated explanations

Example: Complex Code Review

Traditional approach (~2000 tokens):

Review this code for bugs, security issues, performance problems...

Context-engineered approach (~800 tokens):

## Review: auth.py

### Focus Areas
1. Security (OWASP Top 10)
2. Error handling
3. SQL injection

### Output
- Issues: severity + line number
- Fixes: specific code suggestions

[code]

Result: Same quality, 60% fewer tokens.

Best Practices

1. Start Simple: Add complexity only when needed
2. Measure Impact: Track quality improvements
3. Iterate: Refine patterns based on results
4. Document: Keep notes on what works
5. Share: Contribute successful patterns

Integration

Works with:

• Claude Code
• Cursor
• VS Code + Continue
• Any LLM-based tool

Creating Custom Patterns

# Pattern: [Name]

## When to Use
[Trigger conditions]

## Process
[Step-by-step]

## Example
[Concrete example]

## Metrics
[How to measure success]