J-E Translation Review Prompt Chain by an Outbound-Patent Engineer

🔗What Is a Prompt Chain?

A prompt chain is a system of multiple prompts (instructions to AI) linked in stages. This template divides Japanese-to-English translation review into five stages:

1. Basic information confirmation
2. Preliminary review (source text issues, etc.)
3. Basic review (important translation quality items)
4. Detailed review (individual element checks)
5. Overall assessment and feedback creation

At each stage, humans verify results and provide additional instructions as needed.

📋Why Templates Are Necessary

1. Systematizing Review Items

Patent translation review involves numerous aspects:

• Technical terminology consistency
• Compliance with destination country requirements
• Semantic consistency with source text
• Appropriateness of legal expressions

A “framework” is needed to systematically verify these elements.

2. Leveraging AI’s Strengths

AI excels at:

• Term extraction and listing
• Formal consistency checking
• Terminology comparison with applicant websites

However, human verification remains essential for technical validity judgments.

3. Foundation for Customization

Each firm has:

• Different priority review items
• Client-specific requirements
• Technical field specialties

A customizable template effectively addresses these variations.

🔍Common Issues Discoverable Through This Template

This template is designed to help discover issues such as:

Terminology Inconsistencies

Mixed use of “support member” and “supporting member” for the same component can confuse readers and indicates lack of consistency.

Source Text Issues

Translation processes sometimes reveal problems in the original text, such as figure reference errors or duplicate paragraph numbers. While these cannot be corrected in translation, they can be reported as added value to clients.

Technical Validity

The template can identify technically impossible expressions (such as reversed relationships between materials and products) even when grammatically correct.

📊Prompt Chain Template Overview

20+ Item Structure

This template comprises 20+ checkpoints. Key items include:

1. Section heading check – Alignment with destination country standards
2. Product liability considerations (optional) – Avoiding risk-suggesting expressions
3. Technical term extraction and verification – Completeness of terms with drawing reference numerals
4. Applicant-specific terminology check – Alignment with company’s official terminology
5. Semantic consistency between source and translation – Accuracy of subject-object relationships
6. Translation completeness – Checking for omissions and mistranslations
7. Technical validity – Logical process confirmation

(The template containing all items is available via the download link at the end of this article)

Step-by-Step Approach

Rather than processing all items at once, verify in stages:

[Stage 0: Basic Information Confirmation]
→ User confirmation
[Stage 1: Preliminary Review] (Items 1-2)
→ User confirmation
[Stage 2: Basic Review] (Items 3-7)
→ User confirmation
[Stage 3: Detailed Review] (Items 8-19)
→ User confirmation
[Stage 4: Overall Assessment] (Item 20)

🛠️Customization Methods

1. Removing Unnecessary Items

Example: For US filing specialists

• Remove Europe/China-specific requirement checks
• Elaborate on US-specific requirements

2. Adding Custom Items

Example: For pharmaceutical field

21. Pharmaceutical Terminology Check
"Verify that expressions related to drug dosage and administration comply with FDA guidelines. Pay particular attention to:
- Proper use of dosage/dose
- Standard expressions for administration route
- Clinical trial phase descriptions"

3. Adjusting Prompts

Modify existing prompts for your practice:

Before:

List technical terms in Japanese and their English translations in tab-delimited format.

After (Detailed version):

List technical terms in Japanese and their English translations in tab-delimited format.
Categorize them as follows:
1. Equipment/Component names
2. Process/Method names
3. Material/Substance names
4. State/Property names

Also, identify any differences from the client's preferred terminology (registered in project knowledge).

"Please customize this prompt chain template for pharmaceutical field applications"

"Create a US filing-only version. Remove Europe and China items"

"We handle many electrical/electronics patents, so please optimize for that field"

User: "We handle many chemical patents, so please create a version emphasizing IUPAC nomenclature checks"

Claude: "Understood. I'll customize for the chemical field:
- Adding IUPAC nomenclature check as an independent item
- Emphasizing chemical formula and reaction notation
- Specializing Item 16 for chemistry
[Generates customized template]"

5. Creating from Scratch (Also Simple)

“But how do you create a prompt chain template in the first place?” you might wonder.

Actually, no special knowledge is required. I created mine from scratch using this method:

Step-by-Step Creation Process

1. Conduct translation review in regular chat
   • Perform actual translation review work through dialogue with Claude
   • Proceed while confirming step by step
2. Convert procedures to template

   "Please create a prompt chain template from the translation review procedures 
   we performed in this chat that can be used for other projects"

3. Save the template
   • Save the generated markdown template locally
4. Try on next project
   • Register saved template in project knowledge (recommended)
   • Or paste at the beginning of a new chat
   • Use for actual translation review
   • After review completion, ask AI to extract improvement points

     "Looking back at this translation review,
     please list the points that need improvement 
     in the prompt chain template"

5. Improve while using

   "Please update the prompt chain template in project knowledge 
   to reflect the improvements needed from this review"

   💡 Practical Tip
   It’s recommended to perform translation review work and template improvement in separate chats. Doing both in the same chat can quickly hit character limits and cause confusion by mixing different purposes.

6. Refine through repetition
   • Naturally optimized by repeating steps 3-5

In short, just cycle through “Actual work → Template creation → Improvement”.

No special prompt engineering knowledge needed. You can directly convert your daily practice into templates.

💻Implementation with AI Tools (Claude)

Operations Using Project Features

1. Creating Projects
   • Create projects per case or client
   • Register customized prompt chains
2. Registering Required Materials

   Register in project knowledge:
   - Customized prompt chain (.md format)
   - Japanese source text (.md format)
   - English translation (.md format)
   - Client glossaries
   - Past issue lists

3. Starting Review

   After completing the registration of required materials, start with the following command:
   "Please start the Japanese-to-English translation review 
   according to the prompt chain in the project knowledge"

Operational Tips

• Stage-by-stage confirmation: Always have human verification at each stage
• Additional instructions: Ask follow-up questions if AI responses are insufficient
• Result recording: Separately record and accumulate important findings

🔧Integration with Other Tools

Effective Integration with Iro de Check

This template can be effectively integrated with “Iro de Check,” widely used in the patent translation industry.

⚠️Practical Considerations

Appropriate Management of Confidential Information

When using AI tools, please note:

• For content containing unpublished technical information or trade secrets, review each AI service’s terms of use and data handling policies
• For particularly sensitive information such as pre-filing invention content, make decisions at your own discretion

Human Judgment Is Final

• Use AI suggestions as support tools only, with final decisions always made by humans with professional expertise
• Confirmation by specialists in the relevant field and jurisdiction is essential for technical validity and legal judgments

When Source Text Issues Are Found

Translation review processes may reveal source text issues (figure reference errors, duplicate paragraph numbers, etc.). While these cannot be corrected in translation, reporting them as “source text issues” provides added value.

📚Real-World Template Applications

Case 1: Improving Terminology Consistency

One user improved the Item 3 (technical term extraction) prompt:

Improved version:
"Extract technical terms and present in the following format:
Japanese | Translation 1 | Translation 2 (if any) | Recommended | Reason

When multiple translations exist for the same term, 
indicate recommendations based on technical field conventions."

This enabled efficient discovery and correction of terminology inconsistencies.

Case 2: Applicant-Specific Checks

For firms with many cases from specific clients:

Additional item:
"For the following key terms used in [Client Name]'s US patents 
over the past 5 years, verify usage in this translation:
- control unit → Translation of 制御部
- processing module → Translation of 処理モジュール
[Additional client-specific key term list]"

📊Setting Realistic Expectations

What This Template Can Do

• Systematize review work
• Remind about easily overlooked items
• Streamline mechanical verification tasks
• Provide a starting point for customization

Where Human Judgment Is Essential

• Final technical validity determinations
• Context-dependent term selection
• Reflecting client intentions
• Legal risk assessment

💯About Overall Assessment: For Constructive Quality Management

The template includes a 100-point overall assessment. Here’s an important clarification:

This evaluates “translation quality,” not “translator performance.”

Why We Include Numerical Scoring

• To ensure consistency in internal quality management
• To clarify improvement points
• To understand difficulty levels across projects

Recommended Usage

1. Understanding quality trends internally
2. Discovering challenges in specific fields
3. Preparing for constructive dialogue with translators

What matters most is the partnership with translators to achieve quality improvement together.

📈Continuous Improvement

Building Feedback Loops

1. Accumulating Usage Records
   • Which items were effective
   • What customizations were made
   • What problems AI missed
2. Regular Reviews
   • Monthly evaluation of prompt effectiveness
   • Sharing successful cases across teams
   • Adding new review items
3. Improvement Through AI Dialogue

   "How can this prompt be made more effective?"
   
   "Please suggest review items to prevent 
   [specific mistranslation example]"

🎯Conclusion

Japanese-to-English patent translation review is complex and multifaceted. The prompt chain template introduced in this article provides a “framework” for dealing with that complexity.

The key is using this as a starting point and evolving it to match your practice. While AI isn’t a perfect checker, combined with human expertise, it becomes a powerful support tool.

Please use this template as a foundation to build your own unique prompt chain. We hope your experience leads to further improvements.

🚀Future Developments: Application to Office Action Responses

We are currently exploring ways to apply this prompt chain template approach to foreign patent prosecution (office action responses).

Anticipated Challenges and Solutions

While there are challenges in applying this to prosecution work, we believe they are solvable:

The “Practice → Template → Improvement” cycle developed for translation review shows potential for application to other patent practices.

🔬Reference: A Practical Approach Based on Understanding AI Limitations

Structural Limitations of AI Discovered Through Actual Use

Using this template in practice revealed that AI has the following structural limitations. These are constraints of current AI technology, and understanding them properly enables more effective utilization.

1. Narrow Interpretation of Limited Instructions

Core Issue: When AI receives limiting instructions like “major” or “representative,” it judges that requirements are met with minimal verification. While humans would think “let’s check broadly just to be sure,” AI lacks this judgment.

2. Inability to Say “I Don’t Know”

Core Issue: Even when uncertain, AI generates responses based on speculation. After giving definitive answers, it may turn out that AI didn’t understand from the beginning.

3. Positive Reporting Bias

Core Issue: When the overall impression is favorable, AI tends to conclude “generally correct” and neglects searching for localized serious errors. It tends to give answers that would please the operator.

4. Baseless Weighting and Prioritization

Core Issue: AI may arbitrarily judge what is “important,” “main,” or “priority.” When asked “what else besides the main items,” it cannot answer, and when asked for the basis of prioritization, it becomes vague.

5. Inability to Judge Appropriate Detail Level (Granularity)

Core Issue: AI doesn’t understand the human concept of “just right.” When asked to be concise, it removes important information; when asked to be detailed, it includes unnecessary information. It tends toward 0 or 100 extremes.

6. Unawareness of Document Structure

Core Issue: Information often concentrates in specific sections of patent specifications, but unless explicitly instructed, AI doesn’t consider structural characteristics of documents and samples evenly from the entire text.

7. Superficial Understanding of Technical Context

Core Issue: AI doesn’t automatically connect abstract technical understanding with concrete verification. Even when understanding the overall mechanism, it may miss individual technical errors.

8. Pattern Recognition Failure

Core Issue: AI processes individual incidents independently and doesn’t extract common patterns from multiple incidents unless explicitly instructed. It may not recognize systematic problems even when they exist.

9. Perfunctory Completion of Checklist Items

Core Issue: AI tends to prioritize formally completing given task lists while neglecting substantial quality assurance.

Feasibility Study Results

We examined two approaches to these problems:

Recommended Practical Approach

Specializing in AI’s Strengths

• Technical coherence judgments considering context
• Complex sentence structure analysis
• Terminology comparison with applicant websites
• Verification against technical field conventions

Tasks Best Left to Specialized Tools

• Comprehensive verification of large data sets
• Mechanical matching/mismatching checks
• Consistency maintenance across long documents
• Complete checking of all reference numerals and terms for omissions

Summary

Rather than seeking perfection, it’s important to leverage each tool’s strengths. By understanding AI’s limitations and appropriately combining it with specialized tools like Iro de Check, practical quality management can be achieved.