# Concept Areas - TEMPLATE

## Document Purpose

This file contains learning concept areas and notes supporting a learning objective.

---

## Priority Tier 1: AI/LLM Practical Applications


### Concept Areas for Learning Content

**RAG Systems & Vector Databases**

- How retrieval-augmented generation works end-to-end
- Embedding models: how they work, tradeoffs (OpenAI, Cohere, open-source)
- Chunking strategies: size, overlap, semantic vs. fixed
- Vector database options: Pinecone, Chroma, pgvector, Weaviate, Qdrant
- Retrieval strategies: similarity search, hybrid search, reranking
- When RAG helps vs. when it doesn't (failure modes)

**MCP (Model Context Protocol)**

- Protocol specification and design philosophy
- Building MCP servers (Python, TypeScript)
- Tool definition patterns
- Integration with Claude desktop and other clients
- Security considerations for MCP servers

**Claude Skills**

- Skill architecture and file structure
- SKILL.md design patterns
- How skills interact with computer use capabilities
- Testing and iterating on skills
- When to use skills vs. other approaches

**Tool Usage Mechanics (Deep Dive)**

- How LLMs decide when to use tools vs. respond directly
- Tool call formatting and parsing
- Multi-tool orchestration and sequencing
- Error handling and retry logic
- How tool results are incorporated into context
- Tradeoffs: tool use vs. in-context information

**Claude Code for Web & Companion Infrastructure**

- How the computer use environment works (container, file system, permissions)
- Working directories: /home/claude, /mnt/user-data/uploads, /mnt/user-data/outputs
- What Claude can and cannot see/access
- How file operations work and their limitations
- Caching behavior and what Claude knows about it
- Session persistence and state management
- Network access and restrictions

**Token Economics & Context Management**

- Building on Karpathy's foundation: tokenization edge cases, BPE behavior
- Context window management strategies
- Prompt caching: how it works, when it helps, cost implications
- Cost optimization: batching, caching, model selection
- When to use smaller vs. larger context windows
- Conversation design for efficiency

**Critical Evaluation**

- When to use AI vs. traditional approaches
- Hallucination patterns and detection strategies
- Output validation techniques
- Benchmarking and evaluation frameworks
- Understanding model limitations

**Explicitly excluded** (per discussion):

- Prompt engineering (less important with newer models)
- "AI agent" framing (hype-laden; useful parts covered in tool usage and MCP sections)