🌐 日本語

Priority Saturation — Compliance Rates Degrade When Instructions Accumulate

NOTE

In short: The more instructions you give an LLM simultaneously, the lower the compliance rate for each individual instruction.
"Everything is important" is equivalent to "nothing is important."
This is the scientific foundation behind CLAUDE.md's 200-line limit.

What Is Priority Saturation?

Priority Saturation is the phenomenon where the probability of complying with each individual instruction decreases as the number of simultaneous instructions given to an LLM increases.

Quantitative Evidence

IFScale / ManyIFEval Benchmarks

IFScale measures compliance rates when the number of simultaneous instructions increases incrementally, while ManyIFEval measures compliance rates relative to instruction token volume.

Model	Full Compliance at 10 Instructions	Degradation Pattern	Source
GPT-4o	15%	Exponential (rapid degradation)	IFScale / ManyIFEval
Claude 3.5 Sonnet	44%	Linear (gradual degradation)	IFScale / ManyIFEval
o3, Gemini 2.5 Pro	High	Threshold (maintained ~150 instructions, then sharp drop)	IFScale

Three Degradation Patterns

1. Threshold Pattern (o3, Gemini 2.5 Pro): Nearly perfect up to ~150 instructions, then sharp decline

2. Linear Pattern (GPT-4.1, Claude Sonnet 4): Gradual degradation proportional to instruction count

3. Exponential Pattern (GPT-4o, LLaMA Scout): Rapid degradation even with small instruction counts

Critical Degradation Point: ~3,000 Tokens

ManyIFEval confirmed that inference performance begins to degrade at approximately 3,000 tokens of instruction volume. This is a fundamental constraint that cannot be improved by prompt engineering techniques (such as Chain-of-Thought).

Why 200 Lines?

The 200-line limit in CLAUDE.md is a design decision grounded in this research:

• 200 lines ≈ approximately 2,000–3,000 tokens
• This aligns with the degradation threshold identified by ManyIFEval
• Staying within 200 lines maintains approximately 30–40 active instructions
• This preserves individual instruction compliance rates at practical levels

Impact on Coding

• Cramming all rules into CLAUDE.md means critical rules (type safety, testing strategy) are ignored with the same probability as trivial ones (indentation width)
• Giving 10 review criteria simultaneously in code review results in more than half being overlooked
• Test coverage verification becomes less rigorous the more check items you add

Mitigation Strategies in Claude Code

The following are representative examples in Claude Code.

Strategy	Mechanism	Why It Works
CLAUDE.md 200-line limit	Limit resident instructions	Keeps simultaneously active instructions below degradation threshold
.claude/rules/	Conditional injection	Distributes instructions, reducing simultaneous count
Skills	On-demand loading	Load task-specific instructions only when needed
Hooks	Out-of-context enforcement	Exclude mechanically verifiable rules from context budget
Start Small principle	Add after observing failures	Prevent accumulation of unnecessary rules

Relationship to Other Structural Problems

Priority Saturation compounds with the following issues:

• Context Rot: As context length increases, instruction effectiveness further degrades
• Lost in the Middle: Instructions placed in the middle are ignored both due to saturation and positional effects
• Prompt Sensitivity: With more instructions, attention spreads thinner, making outputs more susceptible to phrasing variations
• Hallucination: Missing compliance constraints leads to increased output inaccuracy

This constraint is not unique to Claude

As simultaneous instructions increase, compliance with each instruction falls. The degradation curve differs by model. That instruction density has a ceiling is shared.

How it shows up elsewhere:

• Packing rules into a system prompt or custom instructions makes important rules and trivial ones equally ignored
• Listing ten review criteria at once causes more than half to be missed
• What works is not "write everything." It is keeping the number of simultaneously active instructions down

Other tools do not necessarily ship features at the same granularity. Product-independent principles are extracted in Part 11: Cross-LLM Principles.

References

• Jaroslawicz, D., Whiting, B., Shah, P., & Maamari, K. (2025). "How Many Instructions Can LLMs Follow at Once?" Distyl AI. arXiv:2507.11538 — IFScale benchmark measuring compliance degradation across 10–500 instruction densities
• ManyIFEval (2025) — Compliance evaluation under high instruction counts, showing marked degradation around 3,000 tokens

---

Next: Hallucination

Discussion: #10 Priority Saturation