Loading...
🔍
Self-Critique(SC)
Agent evaluates its own outputs through systematic analysis and generates targeted improvements
Complexity: mediumReflection
🎯 30-Second Overview
Pattern: Single agent evaluates and iteratively improves its own outputs
Why: Catches errors, improves quality, ensures completeness without external validation
Key Insight: Generate → Critique → Revise → Repeat until quality threshold met
⚡ Quick Implementation
1Generate:Initial response to task
2Critique:Systematic quality evaluation
3Identify:Specific issues & improvements
4Revise:Address identified weaknesses
5Iterate:Repeat until quality threshold
Example: generate(task) → critique(output) → revise(issues) → validate() → final_output
📋 Do's & Don'ts
✅Define clear, objective critique criteria for the task
✅Use structured evaluation templates (accuracy, completeness, clarity)
✅Set maximum iteration limits (typically 2-3 cycles)
✅Track quality metrics to measure improvement
✅Document critique reasoning for transparency
❌Allow infinite revision loops without stopping criteria
❌Use vague criteria like "make it better"
❌Over-critique acceptable responses
❌Ignore computational cost vs quality gain tradeoffs
❌Apply same critique template to all task types
🚦 When to Use
Use When
- • High-stakes content generation
- • Complex reasoning tasks
- • Quality > speed requirements
- • Iterative improvement needed
Avoid When
- • Real-time applications
- • Simple/straightforward tasks
- • Limited compute budget
- • Subjective quality domains
📊 Key Metrics
Quality Gain
% improvement from v1 to final
Iteration Count
Average cycles to convergence
Critique Accuracy
Valid issues / Total critiques
Token Efficiency
Quality gain / Extra tokens
Time to Converge
Total processing time
Error Reduction
% decrease in mistakes
💡 Top Use Cases
Academic Writing: Draft → Critique logic/evidence → Strengthen arguments → Polish
Code Generation: Initial code → Identify bugs/inefficiencies → Fix → Optimize
Technical Docs: First version → Check accuracy/completeness → Fill gaps → Clarify
Analysis Reports: Raw analysis → Verify conclusions → Add context → Refine insights
Creative Content: Draft → Evaluate engagement/clarity → Enhance → Final edit
References & Further Reading
Deepen your understanding with these curated resources
Contribute to this collection
Know a great resource? Submit a pull request to add it.
🔍
Self-Critique(SC)
Agent evaluates its own outputs through systematic analysis and generates targeted improvements
Complexity: mediumReflection
🎯 30-Second Overview
Pattern: Single agent evaluates and iteratively improves its own outputs
Why: Catches errors, improves quality, ensures completeness without external validation
Key Insight: Generate → Critique → Revise → Repeat until quality threshold met
⚡ Quick Implementation
1Generate:Initial response to task
2Critique:Systematic quality evaluation
3Identify:Specific issues & improvements
4Revise:Address identified weaknesses
5Iterate:Repeat until quality threshold
Example: generate(task) → critique(output) → revise(issues) → validate() → final_output
📋 Do's & Don'ts
✅Define clear, objective critique criteria for the task
✅Use structured evaluation templates (accuracy, completeness, clarity)
✅Set maximum iteration limits (typically 2-3 cycles)
✅Track quality metrics to measure improvement
✅Document critique reasoning for transparency
❌Allow infinite revision loops without stopping criteria
❌Use vague criteria like "make it better"
❌Over-critique acceptable responses
❌Ignore computational cost vs quality gain tradeoffs
❌Apply same critique template to all task types
🚦 When to Use
Use When
- • High-stakes content generation
- • Complex reasoning tasks
- • Quality > speed requirements
- • Iterative improvement needed
Avoid When
- • Real-time applications
- • Simple/straightforward tasks
- • Limited compute budget
- • Subjective quality domains
📊 Key Metrics
Quality Gain
% improvement from v1 to final
Iteration Count
Average cycles to convergence
Critique Accuracy
Valid issues / Total critiques
Token Efficiency
Quality gain / Extra tokens
Time to Converge
Total processing time
Error Reduction
% decrease in mistakes
💡 Top Use Cases
Academic Writing: Draft → Critique logic/evidence → Strengthen arguments → Polish
Code Generation: Initial code → Identify bugs/inefficiencies → Fix → Optimize
Technical Docs: First version → Check accuracy/completeness → Fill gaps → Clarify
Analysis Reports: Raw analysis → Verify conclusions → Add context → Refine insights
Creative Content: Draft → Evaluate engagement/clarity → Enhance → Final edit
References & Further Reading
Deepen your understanding with these curated resources
Contribute to this collection
Know a great resource? Submit a pull request to add it.