🌐 日本語

Part 1: Understanding the Structural Constraints of LLMs

NOTE

LLMs are not omnipotent. Transformer-based models have structural constraints that come from input length and how attention is allocated.
This Part defines those constraints. The subject is the constraints themselves, not a product setup guide.

Why You Need to Know About Structural Problems

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

Claude Code's configuration (CLAUDE.md, rules/, skills/, hooks, and so on) is a design response to structural problems in LLMs. Claude Code is the subject because it is a representative example that can be described in detail and accurately today.

The principles here apply directly to Cursor, Cline, or plain prompt design. The same constraints appear. The same way of thinking applies.

The destination is Part 11: Cross-LLM Principles. This Part defines the constraints. Parts 2 onward show countermeasures in the representative example. Part 11 extracts what does not depend on the product.

Concrete examples in Claude Code:

• CLAUDE.md's 200-line limit → countermeasure for Priority Saturation
• .claude/rules/ conditional injection → countermeasure for Lost in the Middle
• Hooks for mechanical verification → countermeasure for Hallucination

To understand why configurations are designed this way (the Why), you first need to understand what problems LLMs have.

The 8 Structural Problems

LLMs have the following 8 structural problems. These are not "bugs" — they are unavoidable constraints arising from the Transformer architecture and training process.

Context-Related (Problems that worsen as input grows)

Problem	In a Nutshell	Details
Context Rot	Output quality degrades as tokens increase	Even with 200K capacity, degradation begins at just 50K
Lost in the Middle	Information in the middle of context is ignored	Attention concentrates on beginning and end, with over 30% accuracy loss in the middle
Priority Saturation	Overall compliance drops with too many instructions	With 10 simultaneous instructions, GPT-4o shows 15% and Claude Sonnet 44% compliance

Output-Related (Problems with generation reliability)

Problem	In a Nutshell	Details
Hallucination	Generates content that contradicts facts	Mathematically proven to be "impossible to reduce to zero"
Sycophancy	Agrees with users at the expense of accuracy	A side effect of RLHF. Average 58% compliance rate across all models
Knowledge Boundary	Cannot say "I don't know" for out-of-scope questions	No reward for "I don't know" in the training objective function

Input Sensitivity (Problems dependent on prompt phrasing)

Problem	In a Nutshell	Details
Prompt Sensitivity	Results vary significantly by phrasing	Up to 76 accuracy points difference for the same meaning

Temporal (Problems that worsen as conversations grow longer)

Problem	In a Nutshell	Details
Instruction Decay	Rules are forgotten in long conversations	A compound result of the above 7 problems. Average 39% performance degradation in multi-turn

Relationships Between Problems

These problems do not exist in isolation — they amplify each other. The diagram below visualizes how the 8 structural problems cascade and reinforce one another.

3 Major Cascades:

1. Spatial Degradation: Context Rot → Lost in the Middle → Priority Saturation (accelerates as context grows)
2. Reliability Collapse: Knowledge Boundary → Hallucination ↔ Sycophancy (feedback loop)
3. Temporal Compound: All 7 problems → Instruction Decay (everything converges in multi-turn)

Structural Problems × Claude Code Countermeasures Map

LLMs have 8 structural problems. The countermeasures in the table below are representative examples in Claude Code. From Part 2 onward, we examine how each feature addresses these problems. Product-independent principles are extracted in Part 11.

Structural Problem	Overview	Primary Countermeasures (Claude Code)	Related Docs
Context Rot	Output quality degrades as tokens increase	/compact, /clear, context budget management	Part 2, 3, 5, 6, 8
Lost in the Middle	Information in the middle of context is ignored	/compact (50% threshold), conditional rules, Agents	Part 2, 4, 5, 8
Priority Saturation	Overall compliance drops with too many instructions	CLAUDE.md 200-line limit, .claude/rules/, Skills	Part 3, 4, 5
Hallucination	Generates factually incorrect content (structurally unavoidable)	Hooks (mechanical verification), test code, MCP	Part 6, 7
Sycophancy	Agrees with users at the expense of accuracy	Cross-model QA (Agents), Hooks, question design	Part 5, 7
Knowledge Boundary	Cannot say "I don't know" for out-of-scope questions	MCP external references, version pinning, specialized Agents	Part 3, 5, 6
Prompt Sensitivity	Results vary significantly by phrasing	CLAUDE.md writing style, Skills description design	Part 3, 5
Instruction Decay	Rules forgotten in long conversations (compound of 7 problems)	/compact, /clear, Hooks, session splitting	Part 7, 8

For the detailed version, see Structural Problems × Claude Code Countermeasures Map (Appendix).

This constraint is not unique to Claude

The eight problems are not defects of a specific product. They arise from Transformer-based models and their training. In any environment that uses cloud LLMs, the same constraints appear, to varying degrees.

How they show up elsewhere:

• As chat history grows, agreements made in the middle stop showing up in later replies
• When a long spec is pasted in one shot, requirements in the middle drop out
• When too many rules are listed at once, none of them is followed well

The Claude Code countermeasure table is a representative set of concrete responses. Other tools do not necessarily ship features at the same granularity. What transfers is the way of thinking: keep input short, do not bury important instructions, put verification outside the model. The details are collected in Part 11: Cross-LLM Principles.

---

Next: Part 2: Understanding the Context Window