Ai Prompt Engineering

What it does

This skill provides operational guidance for crafting optimized prompts for various AI models to maximize output quality and achieve desired results. It focuses on practical patterns, checklists, and templates for production-ready prompts across tasks like RAG workflows, agent orchestration, structured outputs, and multi-step planning. The skill includes optimizations specifically for Claude Code + Codex CLI and emphasizes best practices such as versioned prompts, explicit output contracts, regression tests, and safety threat modeling (aligned with OWASP LLM Top 10).

When to use it

• When building API integrations requiring machine-parseable outputs like JSON.
• For extracting specific fields from data sources like forms or invoices.
• To leverage retrieved context from knowledge bases or documentation search using RAG workflows.
• When complex decisions require internal reasoning without verbose explanations.
• To orchestrate multi-step workflows involving API calls and tool usage through agent planning.

Key capabilities

• Structured Output: Creates prompts for generating JSON outputs with schema validation.
• Deterministic Extractor: Designs prompts to extract data precisely, handling missing values appropriately.
• RAG Workflow: Guides prompt creation for effectively using retrieved context with citation checks.
• Hidden Chain-of-Thought: Enables internal reasoning while only outputting the final answer.
• Tool/Agent Planner: Facilitates planning and executing sequences of tool calls, one per turn.
• Rewrite + Constrain: Allows transforming text while preserving meaning and adhering to specific formatting rules.
• Style Matching: Influences the output style through prompt formatting.
• Domain-Specific Patterns: Offers specialized guidance for frontend development, research, and agentic coding.

Example prompts

• "Using the Structured Output pattern, create a prompt that extracts customer data from this text and outputs it as JSON according to this schema: [schema details]."
• "Apply the RAG Workflow pattern to build a prompt that searches our documentation for information on [topic] and includes citations."
• "Using the Tool/Agent Planner pattern, design a prompt that first checks the weather in London, then translates it into Spanish."

Tips & gotchas

• Start with a template from the assets/ directory and fill in the TASK, INPUT, RULES, and OUTPUT FORMAT.
• Prioritize "brief justification" over requesting chain-of-thought reasoning.
• When using private reasoning patterns, instruct the AI to “think internally; output only the final answer.”