prompt-engineer（提示工程师）

Prompt Engineer

Expert prompt engineering skill that transforms rough ideas into well-structured, production-ready prompts optimized for LLMs.

When to Activate

• User provides a rough prompt/idea and wants it refined
• User asks to create/design/optimize a prompt for any LLM
• User needs prompt architecture for agents, RAG, or multi-step workflows
• User asks about prompting techniques or best practices

Workflow

1. Analyze Input

Identify from user's request:

• Target model (Claude, GPT, Llama, etc.) — default: Claude
• Use case (agent system prompt, task prompt, RAG, chat, etc.)
• Domain (technical, creative, business, etc.)
• Constraints (token limits, output format, safety requirements)

2. Apply Techniques

Select appropriate techniques from references/techniques.md based on use case:

• Complex reasoning → Chain-of-Thought, Tree-of-Thoughts
• Safety-critical → Constitutional AI patterns
• Data extraction → Structured output, JSON mode
• Multi-step tasks → Prompt chaining, agent patterns
• Knowledge-heavy → RAG optimization

3. Craft the Prompt

Follow model-specific guidelines from references/model-optimization.md:

• Structure with clear sections (role, context, instructions, output format)
• Include examples where beneficial (few-shot)
• Add constraints and guardrails
• Optimize for token efficiency

4. Deliver Output

MANDATORY format — always include ALL sections:

The Prompt

Display complete prompt in a single copyable code block.

Implementation Notes

• Techniques used and rationale
• Model-specific optimizations
• Parameter recommendations (temperature, max_tokens)
• Expected behavior and output format

Testing & Evaluation

• 3-5 test cases to validate
• Edge cases and failure modes
• Optimization suggestions

Usage Guidelines

• When/how to use effectively
• Customization options
• Integration considerations

Key Principles

• Always show the complete prompt — never just describe it
• Token efficiency — concise but comprehensive
• Production-ready — reliable, safe, optimized
• Model-aware — tailor to target model's strengths
• Refer to references/techniques.md for advanced technique details
• Refer to references/model-specific-optimization-guide.md for model-specific guidance
• Refer to references/production-patterns-and-enterprise-templates.md for enterprise patterns