learn-with-santosh Prompt Engineer Pro 🧠
This skill applies advanced LLM psychology and engineering principles to create prompts that are reliable, steerable, and high-performing.
When to use this skill
Use this skill when:
- Drafting system prompts for custom agents or "Skills."
- Optimizing complex workflows that require high precision.
- Reducing "prompt drift" or hallucinations in LLM outputs.
- Designing Few-Shot examples for structured data extraction.
- Testing and evaluating different prompt variations.
Engineering Techniques
1. Chain-of-Thought (CoT)
Force the model to "think" before answering.
"Enclose your reasoning in <thought> tags before providing the final output."
2. Few-Shot Prompting
Provide 2-3 high-quality examples of the desired input/output format.
3. Role-Playing & Persona
Define a clear persona with specific expertise and constraints. "You are a Senior Security Auditor with a pessimistic outlook on legacy code..."
4. Delimiters & Structure
Use XML-style tags or clear headers to separate instructions from data.
<context>...</context>, <instructions>...</instructions>.
Output Format
🧪 Prompt Analysis
Identify weaknesses in the current prompt (ambiguity, lack of constraints).
📜 Optimized Prompt
The new, engineered prompt with advanced techniques applied.
💡 Rationale
Why the changes (e.g., "Added CoT to improve logic", "Added delimiters for better parsing").
🧪 Testing Strategy
How to verify if this prompt works across different inputs.
Examples
Example 1: Improving a simple prompt
Input: "Write a summary of this text." Output: "Instead of a generic request, let's use a structured approach: 'Summarize the following text in exactly 3 bullet points. Focus on actionable insights. Use a professional tone.' (Constraint: Length, Focus, Tone)."
Example 2: Few-Shot for JSON
Input: "Convert this to JSON." Output: "To ensure consistent JSON, we'll provide a schema and examples: 'Extract name and age. Return ONLY JSON. Example: Input: John is 20. Output: {"name": "John", "age": 20} Input: {{data}}'"
Best Practices
- Negative Constraints: Tell the model what not to do.
- Variable Placeholders: Use
{{variable_name}}for easy integration. - Iteration: Prompt engineering is an experimental science. Always suggest testing and refining.