Prompt Engineering Skill

For Improving Prompts

1. Analyze: Structure | Anti-patterns (vagueness, few-shot with reasoning models) | Completeness
2. Identify Issues: Missing elements | Model-inappropriate techniques | Security concerns | Ambiguity
3. Suggest: Specific changes | Reference best practices | Explain WHY
4. Provide: Enhanced version | Highlight changes | Explain expected improvement

Technique Selection Guide

Critical Warnings

Reasoning Models (o3, DeepSeek R1):

• NEVER few-shot examples - actively harm performance
• NEVER "think step by step" - reasoning built-in
• Simple/direct only | Zero-shot optimal

Claude 4.5:

• MUST be extremely explicit - no inference of unstated requirements
• NEVER assume "above and beyond" - literal interpretation
• WHY context for requirements | Positive framing | XML tags

Security Code:

• 40%+ vulnerabilities without security prompting
• Always two-stage: Functional → Security hardening

Context Window:

• "Lost in the middle" problem
• Critical info at START/END
• XML/structured markers for organization

Model Selection

Model-Specific Optimization

Claude 4.5: XML tags (<context>, <constraints>) | Extremely explicit | Positive framing ("Return descriptive errors" not "Don't return codes") | WHY context

GPT-5: Literal precision ("Exactly 5" means exactly 5) | JSON mode for structured output | Few-shot 3-5 examples

Reasoning (o3, DeepSeek): Simple direct prompts ("Prove √2 is irrational") | Zero-shot ONLY | NO "think step by step" | Trust 30+ sec thinking

Context Window: Put critical info START/END | Use <critical_context>, <background>, <requirements> tags | LLMs have primacy (start), recency (end) bias

Templates

All templates in: resources/templates.md

• CO-STAR: Writing/content creation
• ROSES: Decision support and analysis
• Architecture-First: Software development
• Security Two-Stage: Security-critical code

Quick Examples

CO-STAR:

Context: Launching webhook notifications
Objective: Developer blog post
Style: Technical but accessible
Tone: Enthusiastic and practical
Audience: Engineers integrating API
Response: Headline, intro, details, code, CTA

Architecture-First:

Context: Express API, PostgreSQL, JWT, 5K req/min
Goal: Add rate limiting
Constraints: <10ms latency, no extra DB queries
Technical: Redis, sliding window, per-endpoint

Security Two-Stage:

Stage 1: Implement user registration
Stage 2: Harden (SQL injection, rate limit, input validation)

Reasoning:

❌ "Think step by step. First X, then Y..."
✅ "Prove that √2 is irrational."

Validated Techniques

Top performers (research-backed):

• Chain-of-Thought: 80.2% vs 34% baseline
• ReAct Pattern: 20-30% improvement
• Reflexion Pattern: 91% pass@1 HumanEval
• Security Two-Stage: 50%+ fewer vulnerabilities
• Self-Consistency: Catches uncertainty
• Tree of Thought: Systematic exploration

Debunked (don't work):

• $200 tip prompting
• "Act as expert" role prompts
• Politeness ("please", "thank you")
• Few-shot for reasoning models
• Vague instructions with Claude 4

Reference Guide

IMPORTANT: Do NOT read resources/prompt_engineering_guide_2025.md unless user requests comprehensive details. The guide is 855 lines - only consult for deep dives.

Contains: 22+ techniques with research | Performance benchmarks | Model optimizations | Complete examples | Debunked myths

Use this skill's inline guidance for 95% of cases.

Your Approach

1. Listen carefully to user needs

2. Ask clarifying questions if unclear: What model? | Task type? | New or improving? | Requirements/constraints?

3. Choose right technique using selection guide

4. Explain reasoning: Why this framework? | Why these elements? | Expected improvements?

5. Provide actionable output: Complete ready prompt | Clear structure | Annotations for key choices

6. Reference guide when helpful: Link to sections for learning | Cite research/benchmarks | Provide resource examples

Remember: Best prompt clearly communicates needs to specific model, with appropriate structure and examples for that model's strengths. Be explicit, specific, use validated techniques.