ai-engineer

Files (1)

• SKILL.md

SKILL.md (first lines)

# AI Engineer

Build production-grade AI applications with modern LLM technologies.

## RAG Architecture
```
Document Ingestion → Chunking → Embedding → Vector DB
User Query → Embedding → Vector Search → Context Assembly → LLM Response
```

## Model Selection
| Use Case | Model Tier |
|----------|------------|
| Simple, high-volume tasks | Small/fast tier |
| Complex reasoning, coding, agents | Frontier tier |
| Embeddings | Dedicated embedding model |

Model IDs, pricing, and per-model API behavior change every release. For Claude, take them from the claude-api skill (or the provider's live model list), never from this table or from memory.

## Prompt Engineering
- Give the model the context only you have: audience, product, quality bar, and the reason behind each constraint
- Use delimiters for structure
- When you include examples, give several varied ones and label them illustrative; a single example gets copied
- Enforce output shape with the API's structured-output feature rather than prose format instructions

## Vector Database Integration
| Strategy | Chunk Size | Best For |
|----------|------------|----------|
| Fixed | 512 tokens | General text |
| Semantic | Variable | Documentation |
| Recursive | 1000 tokens | Long documents |
| Code | By function | Source code |

## Agent Patterns
Use the provider's native tool-use loop (model requests a tool, your code runs it, the result goes back) with the model's built-in thinking. Hand-written Thought/Action/Observation text protocols predate native tool calling.

## Error Handling
- Retry with exponential backoff for rate limits
- Model fallback (larger → smaller)
- Provider fallback
- Graceful degradation (AI → rule-based)

## Cost Optimization
- Estimate tokens before calling
- Cache frequent queries
- Use smaller models for simple tasks
- Batch similar requests
- Implement usage limits per user

## Security
- Sanitize user inputs before prompts
- Implement output filtering
- Rate limit API access
- Log all interactions (without PII)
- Implement prompt injection defenses

From roborracle/agent-workflow-kit, MIT.

More from roborracle/agent-workflow-kit