🌐 日本語
Structural Problems × Claude Code Countermeasures Map (Detailed)
NOTE
Detailed correspondence between the 8 structural problems and Claude Code features.
Countermeasure Map
Context Rot (Quality Degradation with More Tokens)
| Countermeasure | Category | Effect |
|---|---|---|
/compact | Session Management | Compress conversation history to reduce token count |
/clear | Session Management | Reset session to start with fresh context |
| CLAUDE.md 200-line limit | Resident Context | Minimize fixed cost |
.claude/rules/ | Conditional Context | Prevent unnecessary rule injection |
| Skills | On-Demand | Deploy only when needed |
| Agents | On-Demand | Avoid root cause with independent context |
| MCP Tool Search | Tool Definition | Lazy-load tool definitions |
| Code Intelligence (LSP) | Code World | Retrieve symbols precisely without loading whole files |
| Agent Teams | Multi-Session | Parallel decomposition — no single session ever grows large |
Lost in the Middle (Information Loss in Middle of Context)
| Countermeasure | Category | Effect |
|---|---|---|
/compact (50% threshold) | Session Management | Compress before U-curve collapse |
.claude/rules/ | Conditional Context | Inject only necessary rules at high-attention positions |
| Agents | On-Demand | Avoid middle problem with fresh context |
| Skills | On-Demand | Inject near end for high-attention placement |
| Agent Teams | Multi-Session | Each session's history stays short; no "middle" forms |
Priority Saturation (Reduced Compliance with Too Many Instructions)
| Countermeasure | Category | Effect |
|---|---|---|
| CLAUDE.md 200-line limit | Resident Context | Keep simultaneous active instructions below degradation threshold |
.claude/rules/ | Conditional Context | Distribute instructions conditionally |
| Skills | On-Demand | Load task-specific instructions only when needed |
| Hooks | Runtime | Exclude mechanical rules from context budget |
| Agent Teams | Multi-Session | Split responsibilities across sessions; each carries fewer instructions |
Hallucination (Structurally Unavoidable)
| Countermeasure | Category | Effect |
|---|---|---|
| Hooks (test execution) | Runtime | Compilers and test runners don't hallucinate |
| Cross-Model QA (Agents) | On-Demand | Verification across different models |
| MCP | Tool Definition | Reference external trusted sources |
| Code Intelligence (LSP) | Code World | Symbol-level grounding before code is committed |
| CLAUDE.md | Resident Context | Make constraints and version info explicit |
Sycophancy (Prioritizing Agreement Over Accuracy)
| Countermeasure | Category | Effect |
|---|---|---|
| Agents (Cross-Model QA) | On-Demand | Different models don't share the same conformity bias |
| Hooks | Runtime | Mechanical verification without conformity bias |
| CLAUDE.md (contradiction instruction) | Resident Context | Explicit instruction to "find problems" |
| Test Code | External Validation | Objective facts as a safeguard |
Knowledge Boundary (Inability to Say "I Don't Know")
| Countermeasure | Category | Effect |
|---|---|---|
| MCP (external knowledge reference) | Tool Definition | Extend knowledge boundary externally |
| Code Intelligence (LSP) | Code World | Resolve project-private and post-cutoff symbols via language server |
| CLAUDE.md (explicit version) | Resident Context | Specify "which point in time's knowledge to use" |
| Agents (knowledge separation) | On-Demand | Narrow knowledge domain to reduce boundary crossing probability |
| Test Code | External Validation | Detect outputs that exceed knowledge boundary |
Prompt Sensitivity (Results Vary by Expression)
| Countermeasure | Category | Effect |
|---|---|---|
| CLAUDE.md writing style | Resident Context | Specific and imperative writing removes ambiguity |
| Skills description | On-Demand | Diverse expressions improve invocation accuracy |
.claude/rules/ | Conditional Context | Reduced simultaneous instructions prevent sensitivity degradation |
| Hooks and tests | Runtime | Verification independent of prompt expression |
| Plugins / Marketplaces | Distribution | Distribute verified prompts so each engineer does not retry from scratch |
Instruction Decay (Rule Forgetting in Long Conversations)
| Countermeasure | Category | Effect |
|---|---|---|
/compact | Session Management | Preventive compression before 50% usage |
/clear | Session Management | Reset degradation by splitting sessions |
| Hooks | Runtime | Force execution independent of context |
| Agents | On-Demand | Execute task with fresh context |
| Git commits | External Persistence | Easy rollback of degraded output |
| Plugins / Marketplaces | Distribution | Enforce organizational conventions so they survive turnover and onboarding |
Full Map (Visual)
Visualize the above countermeasure map from the perspective of countermeasure categories. Each diagram shows "which problems this countermeasure addresses."
Session Management — /compact /clear
Resident Context — CLAUDE.md (200-line limit)
Conditional Context — .claude/rules/
On-Demand Context — Skills & Agents
Runtime — Hooks
Tool Definition — MCP & Tool Search
Integrated Full Map — Problem Chain × Countermeasure Placement
The complete picture of how the 8 structural problems interconnect and where Claude Code features intervene.
How to Read:
| Element | Shape | Meaning |
|---|---|---|
| ■ Rectangle nodes (color-coded) | Structural problems (color indicates type) | |
| ⬮ Rounded nodes (blue) | Claude Code countermeasures | |
| Solid → | Problem causes or amplifies another problem | |
| Dotted -.-> | Countermeasure intervenes at this point |
Countermeasure Map (Overview) also available