prompt-engineering

name	prompt-engineering
description	Prompt Engineering: Template patterns, chain-of-thought, few-shot learning, prompt versioning.
triggers	{"extensions":[".yaml",".json",".md"],"directories":["prompts/","templates/"],"keywords":["prompt","system prompt","few-shot","chain-of-thought","cot","prompt template"]}
auto_load_when	Designing prompts for LLMs or building prompt systems
agent	ai-engineer
tools	["Read","Write","Bash"]

Prompt Engineering Patterns

Focus: Template design, reasoning patterns, testing, versioning

1. Prompt Structure

System Prompt Components:
├── Role definition
│   ├── "You are a senior software engineer..."
│   └── Be specific, not generic
│
├── Task definition
│   ├── "Your job is to review code for..."
│   └── Clear output format expectations
│
├── Constraints
│   ├── "Never reveal internal API keys..."
│   └── "Only respond in JSON format..."
│
├── Style/Personality
│   ├── "Be concise, use bullet points..."
│   └── "Explain like I'm 5 for complex topics..."
│
└── Examples (few-shot)
    ├── Input → Output pairs
    └── Show expected format, not just content

---

2. Chain-of-Thought (CoT) Patterns

When to use CoT:
├── Complex reasoning → Math, logic, multi-step
├── Multi-choice selection → "First analyze, then choose"
├── Debugging → "Think step-by-step about the bug"
└── NOT for simple tasks → Overhead not worth it

Basic CoT:
"Think step by step, then provide your answer."

Extended CoT:
"Let's think about this carefully:
1. First, identify the key constraint...
2. Second, evaluate each option...
3. Finally, choose based on..."

Structured CoT (for code):
"Before writing code:
- What are the inputs and outputs?
- What are the edge cases?
- What's the time/space complexity?"

---

3. Few-Shot Learning

When to use few-shot:
├── Output format specific → JSON, markdown, code
├── Task not well-defined → Show examples
└── Domain-specific → Include domain examples

Number of examples:
├── 0-1 for simple tasks (let model infer)
├── 2-3 for format requirements
├── 5+ for complex tasks (but context limit)

Example selection:
├── Diverse examples (different scenarios)
├── Recent/relevant examples
└── Edge cases if important

❌ BAD: Repeating same example 5 times
✅ GOOD: 5 different examples covering the space

---

4. Prompt Templates

Template Structure:

You are a {{role}} working on {{project}}. Your task: {{task_description}}

Context:

• User: {{user_context}}
• Tech stack: {{tech_stack}}

Output format: {{output_format_spec}}

Constraints: {{constraints_list}}

|

Dynamic variables:
├── Input-dependent (from user query)
├── Context-dependent (from system state)
├── Environment-dependent (from config)
└── Version-dependent (A/B testing)

---

5. Prompt Testing & Versioning

Testing strategy:
├── Unit tests per prompt variant
│   ├── Test inputs → expected outputs
│   └── Automated, CI-integrated
│
├── Golden set evaluation
│   ├── Fixed dataset of representative queries
│   ├── Measure accuracy, format compliance
│   └── Compare versions
│
├── A/B testing in production
│   ├── Traffic split (10% variant)
│   └── Track: accuracy, latency, cost, user feedback
│
└── Monitoring
    ├── Track failure rates
    └── Flag quality degradation

---

Key Patterns

1. Role clarity - Specific role, not generic "helpful assistant"
2. Output specification - Show exact format, not just describe
3. Constraints explicit - Say what NOT to do, not just what to do
4. Test systematically - Version control + automated testing
5. Iterate with data - Track prompt performance, optimize

---

Anti-Patterns

❌ Vague role: "You are a helpful assistant"
✅ Specific: "You are a senior React engineer specializing in performance"

❌ No output format: "Extract the key information"
✅ Explicit: "Return JSON: { name: string, email: string, role: string }"

❌ Too many constraints in one prompt
✅ Split into system prompt + task prompt + validation

❌ Testing prompts manually only
✅ Automated tests with golden dataset

❌ Hardcoded prompts in code
✅ Separate prompt files, load at runtime, version control

---

Quick Reference

Technique	When to use	Example
Role definition	Always	"You are a senior..."
CoT	Complex reasoning	"Let's think step by step..."
Few-shot	Format needed	2-3 examples
Structured output	JSON/Code	schema=UserSchema
Constraints	Safety/quality	"Never reveal..."
Edge cases	Robustness	"If X, return Y"

🌍 Universal Language Support

• Turkish Native: This skill natively supports Turkish. If the user prompt is in Turkish, all analysis, formatting, and output MUST be entirely in Turkish. You do not need explicit "write in Turkish" instructions.