llm-application-dev_skill

Overview

This skill explains building applications with large language models, covering prompt engineering, RAG patterns, API integration, streaming, and error handling. It focuses on practical patterns for chatbots, retrieval-augmented features, and safe, cost-aware LLM integration. The content is implementation-oriented and geared for production use.

How this skill works

It describes structured prompt patterns (system/user, few-shot, chain-of-thought) and shows how to call LLM APIs, handle streaming, and implement retries and backoff. For RAG, it covers embedding, vector search, document chunking, and storing embeddings for similarity search. It also outlines operational practices like token management, caching, evaluation, logging, and privacy safeguards.

When to use it

• Building chatbots or conversational agents
• Adding knowledge-grounded answers via retrieval-augmented generation
• Automating tasks that require natural language understanding or generation
• Prototyping and productionizing LLM features with attention to costs and latency
• Integrating streaming responses for improved user experience

Best practices

• Use structured system and user prompts and limit response length to control tokens
• Cache embeddings and frequent query results to reduce cost and latency
• Chunk large documents with overlap to preserve context for RAG searches
• Implement retries with exponential backoff and sensible fallbacks for transient errors
• Validate and sanitize LLM outputs; never trust generated PII without checks
• Log prompts, responses, and metrics for evaluation and safe rollout

Example use cases

• Customer support chatbot that answers only from product docs via RAG
• Sentiment or intent classification using few-shot prompts for edge cases
• Knowledge base Q&A with document chunking and vector search
• Live streaming assistant that progressively renders generated content to users
• Automated report generation combining retrieved facts and structured prompts