Understanding LLMsThrough Claude Code

Understand the structural constraints of LLMs and learn why designs look the way they do. The subject is Claude Code. The principles are not product-specific.

Start Reading

View on GitHub

Part 1: Structural Problems

The 8 limits LLMs cannot overcome by themselves — Context Rot, Lost in the Middle, Hallucination, and more.

Part 2: Context Window

The LLM's thinking space — tokens, context budget, and injection timing.

Part 3: Always-Loaded Context

CLAUDE.md design, hierarchical merging, and CLAUDE.local.md.

Part 4: Conditional Context

Rules and glob patterns — context loaded only when files match.

Part 5: On-Demand Context

Skills and Agents — context loaded only when the LLM asks for it.

Part 6: Tool Context — MCP

Tool definitions as context cost, and how Tool Search avoids it.

Part 7: Runtime Layer

settings.json and Hooks — the layer LLMs never see.

Part 8: Session Management

/compact, /clear, and what to persist across sessions.

Part 9: Code Intelligence

LSP-driven grounding for code symbols — the second pillar alongside MCP.

Part 10: Multi-Session Coordination

Agent Teams as the root-cause fix for Context Rot at scale.

Part 11: Cross-LLM Principles

The destination of the whole work. Extract product-independent principles and apply them to Cursor, Cline, and plain prompt design.

Appendix

Maps, configuration reference, FAQ, plugins, and glossary.

About this site

LLMs are not omnipotent. Transformer-based models have structural constraints: quality falls as input grows, information in the middle is missed, and content that contradicts facts is generated.

This site is a "bookshelf of Why." It is for learning those constraints and the design principles that transfer across prompt and agent design.

The intended reader is a developer who uses cloud LLMs daily. The goal is to understand the constraints and apply them in one's own environment.

Claude Code is the main subject because it is a representative example that can be described in detail and accurately today. The structural constraints and design principles here do not depend on a specific product. The same constraints appear in Cursor, Cline, or plain prompt design. The same way of thinking applies.

The destination is Part 11: Cross-LLM Principles. Parts 1–10 confirm the principles through Claude Code as a representative example. Part 11 then extracts what does not depend on the product.

Sister projects

Three sister projects can be read in order: "Know LLMs → Know Agent Design → Apply to Systems." This site is the "bookshelf of Why."

Phase	Project	Focus
1. Know LLMs	This site	LLM structural constraints and why designs look the way they do (bookshelf of Why)
2. Know Agent Design	ai-agent-architecture	MCP, Skills, and Agent composition with implementation patterns (map of What/How)
3. Apply to Systems	Management-of-software-systems-and-services	Coming soon — system operations in the AI era

TIP

After the Why is clear, read the sister site ai-agent-architecture for concrete implementation patterns (MCP catalog, Skills design, Agent taxonomy, A2A protocol, and so on). Why and How then connect.