Glossary

LLM Structural Problems

Term	Definition
Context Rot	Degradation of output quality as input token count increases
Lost in the Middle	Reduction in attention to information in the middle of context
Priority Saturation	Decrease in compliance with individual instructions as simultaneous instruction count increases
Hallucination	Generation of false content with confidence
Sycophancy	Tendency to agree with users at the expense of accuracy
Knowledge Boundary	Inability of LLMs to recognize the limits of their own knowledge
Prompt Sensitivity	Generation of different outputs from semantically equivalent prompts
Instruction Decay	Decrease in compliance with initial instructions during long conversations

Term	Definition
CLAUDE.md	Persistent context file for documenting project knowledge and conventions
`.claude/rules/`	Rule files conditionally injected via glob patterns
Skills	Task-specific procedures deployed on demand
Agents	Sub-agents executed in independent context windows
Hooks	Context-external processing executed at lifecycle events
MCP	Model Context Protocol. Connection to external tools and services
settings.json	Configuration file for Claude Code runtime (not injected to LLM)
`/compact`	Command to summarize and compress conversation history
`/clear`	Command to reset a session
Tool Search	Mechanism for lazy-loading MCP tool definitions (automatically enabled when context exceeds 10%)

Term	Definition
Context Window	Maximum range of input an LLM can process at once (Claude 4.6 series: 1M tokens)
Token	Smallest unit processed by LLM. In Japanese, approximately 1-2 tokens per character
Transformer	Foundational architecture of modern LLMs. Features self-attention mechanisms
RLHF	Reinforcement Learning from Human Feedback
RoPE	Rotary Position Embedding. Method for encoding positional information
Cross-Model QA	Quality assurance technique for reviewing outputs across different models
U-Curve	Attention distribution pattern in context. High at beginning and end, low in middle