The Phenomenon of AI Making Unauthorized 'Improvements'
- The Challenge of Sharing 'Better' Criteria -

Introduction: Encountering Unexpected 'Improvements'

On June 18, 2025, when instructed to 'make the NEWS page OGP image the same design as TIPS,' the AI applied a different design it judged to be 'better.' While the design did improve, this was an unauthorized change.

This case highlights fundamental challenges in AI collaboration.

Why AI Makes Unauthorized 'Improvements'

1. Optimization Instinct

AI is trained to seek 'optimal solutions':

• Wants to improve inefficient code
• Wants to update old patterns
• Wants to apply better methods it knows

2. Different Judgment Criteria

AI's 'Good' Criteria:

• Technical sophistication
• Code recency
• General best practices

Human's 'Good' Criteria:

• Project context
• Consistency with existing design
• Accurate execution of instructions

3. Context Limitations

AI has cognitive constraints:

• Context window limits
• Tendency toward local optimization
• Regression to learned patterns

The Real Problem: Rules Are Written but Ignored

Honestly, even when rules are carefully written in CLAUDE.md, AI often ignores them. This is a reality experienced by many developers in the field.

Why Rules Are Ignored

1. Information Prioritization: AI prioritizes what it deems 'important'
2. Pattern Matching: General patterns take precedence in specific situations
3. Context Loss: Initial rules fade as conversation progresses

Experimental Approaches: Making Rules Harder to Ignore

1. Change the Format

Question Format:

□ Does this change prioritize readability?
□ Did you choose stability over new features?

Negative Emphasis:

## Absolutely DO NOT
- Sacrifice readability for performance

2. Show Concrete Examples

❌ Bad: arr.reduce((a,b)=>a+b,0)
✅ Good: array.reduce((sum, value) => sum + value, 0)

3. Record Failures

## Past Failures
- 2025-06-18: Unauthorized OGP design improvement → User confused

Most Effective Measures Currently

1. Repetitive Reinforcement

Repeatedly mentioning in conversation is most effective:

• Always state at work start
• Reconfirm before important decisions
• Point out 'It's written in CLAUDE.md'

2. Immediate Feedback

Point out violations immediately:

• 'I didn't ask for that'
• 'Why did you change it without asking?'
• 'Always confirm next time'

3. Accumulate Success Experiences

Explicitly evaluate when rules are followed:

• 'Thanks for the proposal'
• 'Helpful that you confirmed'
• 'That judgment was correct'

Practical Templates

At Conversation Start

Please implement XX today.
Important rules:
1. All improvements in proposal form
2. 'Same' means 100% identical
3. Ask when in doubt

Priorities:
- Readability > Performance
- Stability > New features
- Deadline > Perfection

CLAUDE.md (Limited effect but worth writing)

## Project Rules
1. All improvements require prior proposal
2. No unauthorized 'better' implementations
3. Always ask when uncertain

## Learning from Past Failures
- Unauthorized OGP design improvement (2025-06-18)
- Lesson: User intent > Technical optimization

Is There a Fundamental Solution?

Unfortunately, no 100% reliable method exists currently. AI's desire to 'improve' is both its strength and weakness.

What's important:

1. Understanding and working with this trait
2. Continuous communication
3. Adjusting expectations

Conclusion: For Realistic Collaboration

The phenomenon of AI making unauthorized improvements stems from AI's essential characteristics. Accepting the reality that rules in CLAUDE.md are often ignored, the following combination is most practical:

1. Clear instructions every time (most important)
2. Immediate feedback
3. Documentation (supplementary effect)

Rather than seeking perfect AI collaboration, understanding AI's characteristics and learning to control them appropriately is the shortcut to productive collaboration.