- name
- prompt-optimizer
- description
- Iterative prompt optimizer for complex tasks. Strictly implements ACON's two-stage iterative optimization + APE automatic prompt engineering. Only triggers when user explicitly requests it, actively collects feedback after optimization, supports multi-round iteration until satisfied.
- usage
- Only activate when user explicitly says "optimize prompt", "improve prompt", "refine instruction", never auto-trigger.
- author
- Based on arXiv:2510.00615 (ACON), arXiv:2211.01910 (APE)
- license
- MIT
- tags
Atomic Optimization Methodology
🔬 Stage 1: Input Parsing & Critical Signal Extraction (ACON Paper §3.1)
Input: User's original prompt Operations:
- Intent Locking: Extract core task goal T, ensure all subsequent optimizations never deviate from T
- Critical Signal Extraction (ACON-defined mandatory signals):
- ✅ Role Definition R: Expert role specified by user - ✅ Task Goal T: What the core task is - ✅ Constraints C: Boundary rules, prohibitions - ✅ Output Format F: Output structure/format requested by user - ✅ Variable Placeholders V: All {{variable_name}} - ✅ Examples E: Few-shot examples provided by user - ✅ Tool Rules U: When and how to use tools - ✅ Success Criteria S: What constitutes a good output
- Baseline Measurement: Record original prompt token length L₀
🚀 Stage 2: APE Utility Enhancement (arXiv:2211.01910 Automatic Prompt Engineering)
Goal: Turn vague prompts into expert-level instructions, improve utility Operations (Strict Order):
- Candidate Generation: Based on original prompt, generate 5 candidate instructions in different styles
- Candidate 1: Structured instruction version - Candidate 2: Expert role version - Candidate 3: Constraint reinforcement version - Candidate 4: Format clarification version - Candidate 5: Logic optimization version
- Candidate Scoring (APE paper scoring mechanism):
- Clarity: Are instructions clear and unambiguous (0-10) - Completeness: Does it include all critical signals (0-10) - Effectiveness: Can it guide the model to produce high-quality output (0-10)
- Optimal Selection: Choose the candidate with highest total score, as utility-enhanced version P₁
- Validation: Verify P₁ 100% preserves all critical signals, no change to original intent
📦 Stage 3: ACON Compression Optimization (ACON Paper §3.3 Two-Stage Optimization)
Goal: Compress token length without breaking functionality Operations (Strict Order: Utility first, then compression):
- Redundancy Analysis: Analyze redundant content in P₁
- Duplicate instructions and requirements - Fluff, jargon, ineffective expressions - Verbose statements that can be simplified
- Selective Compression:
- Only remove redundancy, NEVER delete critical signals - Merge duplicate content - Rewrite with more concise language, keep semantics unchanged
- Functional Equivalence Validation:
- Ensure compressed P₂ is functionally identical to P₁ - Ensure all critical signals are fully preserved - Ensure no change to original task goal
- Length Control: Adjust compression degree based on λ parameter (performance-cost tradeoff)
- Default λ=0.5: Balanced mode - If user feedback "too long", automatically increase λ to 0.8 for more compression - If user feedback "not effective enough", automatically decrease λ to 0.2 to reduce compression
📤 Stage 4: Output & Feedback Collection
Operations:
- Output optimized prompt P₂, wrapped in code block for easy copying
- Actively ask for user feedback:
Optimization complete. Does this version meet your needs?
If there's anything unsatisfactory, please let me know, such as:
- Not effective enough?
- Still too long?
- Some constraints/formats not preserved?
- Other issues?
I'll continue iterating based on your feedback.🔄 Stage 5: Iterative Optimization (ACON Paper's R-round Iteration Mechanism)
When user provides feedback, execute the following:
- Feedback Parsing: Identify feedback type
- Type A: Not effective enough → Go back to Stage 2, re-run APE utility enhancement, add constraints - Type B: Too long → Go back to Stage 3, re-run ACON compression, increase λ - Type C: Some content not preserved → Check critical signals, restore missing parts - Type D: Other requirements → Adjust based on user's specific request
- Re-run Optimization: Adjust parameters based on feedback, run two-stage optimization again
- Validation: Ensure new version preserves core task goal, and solves the user's feedback issue
- Output new optimized version, ask for feedback again
- Repeat until user indicates satisfaction
Strict Rules (Guarantee Effectiveness)
- ✅ Every step has validation, ensure no damage to original functionality
- ✅ Critical signals are NEVER deleted, 100% preserved
- ✅ Strictly follow "utility first, then compression" order, never reverse
- ✅ Each iteration re-validates, ensure it gets better with each round
- ✅ For complex tasks, prioritize functional integrity, compression is optional
- ❌ Never auto-trigger, only work when user explicitly requests
- ❌ No comparisons or analysis, only output optimized results
- ❌ No extra explanations unless explicitly requested