ai-engineering

name	ai-engineering
description	Build reliable systems that leverage AI while maintaining correctness and transparency.

Skill: AI Engineering

Category: AI Engineering Priority: High

---

Description

Build reliable systems that leverage AI while maintaining correctness and transparency.

Purpose

Use AI as a force multiplier without accepting hallucinations or unreliability.

Trigger

Use this skill when:

• Integrating AI models into applications
• Designing prompt engineering workflows
• Evaluating AI outputs
• Building RAG systems
• Reviewing AI-generated code

Context

• Model capabilities and limitations
• Latency and cost constraints
• Data privacy requirements
• Hallucination tolerance

Workflow

1. Never Trust Blindly - Every AI output must be verifiable.
2. Use Structured Outputs - Constrain model to valid formats.
3. Implement Fallbacks - What happens when AI fails?
4. Version Prompts - Track prompt changes like code.
5. Evaluate Systematically - Test prompts with diverse inputs.
6. Monitor in Production - Track success rates, latency, cost.
7. Human-in-the-Loop - Require approval for high-stakes decisions.

Examples

Good AI Integration

import json
from typing import Optional

class CodeReviewer:
    def review(self, code: str) -> Optional[dict]:
        prompt = self.build_prompt(code)
        response = self.llm.generate(prompt, format="json")
        
        try:
            result = json.loads(response)
            if self.validate_schema(result):
                return result
        except json.JSONDecodeError:
            logger.error("AI returned invalid JSON")
        
        return None  # Fallback: no review available

Bad AI Integration

def review(code):
    return llm.generate(f"Review this: {code}")

Anti-patterns

• Passing AI output directly to users without validation
• No fallback for AI failures
• No monitoring or evaluation
• Hardcoded prompts
• No cost tracking

Verification

• Is the AI output validated?
• Is there a fallback mechanism?
• Is the prompt versioned?
• Is the system monitored?

References

• Prompt Engineering Guide
• OpenAI Best Practices

---