Tuesday, April 21, 2026

체리피킹은 그만! 대법원 판결로 본 '선행문헌 전체 대비 원칙' 실무 가이드

Legal Commentary · IP Law

진보성 판단에서 선행문헌 전체를 읽는 법

대법원 2019후12094 판결과 선행기술 결합의 실무

주요 준거 판결 대법원 2022. 1. 13. 선고 2019후12094 판결

대표 판결 대법원 2016. 1. 14. 선고 2013후2873·2880(병합) 판결

진보성 일반 판단 틀 대법원 2007후3660 판결 등

핵심 쟁점 선행문헌 전체 대비, 선택적 발췌, 결합 이유, 부정적 교시, 사후적 고찰

1. 핵심 법리: 일부 문구가 아니라 문헌 전체가 무엇을 가르치는가

진보성 판단에서 선행문헌의 특정 문장이나 수치범위가 청구발명과 겹친다는 사실만으로 결론을 내릴 수는 없다. 법원이 확인해야 할 것은 선행문헌 전체를 읽은 통상의 기술자가 출원 당시 실제로 무엇을 합리적으로 인식했는가이다.

대법원 2022. 1. 13. 선고 2019후12094 판결
진보성 부정의 근거가 될 수 있는 일부 기재만이 아니라 그 선행문헌 전체에 의하여 통상의 기술자가 합리적으로 인식할 수 있는 사항을 기초로 대비·판단하여야 한다.
이 판결은 선행기술의 범위와 내용, 출원발명과의 차이, 통상의 기술자의 기술수준을 증거에 기초하여 파악하고, 출원발명을 이미 알고 있다는 전제에서 사후적으로 판단해서는 안 된다는 일반원칙도 함께 확인하였다.
대법원 2016. 1. 14. 선고 2013후2873·2880(병합) 판결
선행문헌 전체에 의하여 통상의 기술자가 합리적으로 인식할 수 있는 사항을 기초로 대비·판단하여야 하고, 일부 기재와 배치되거나 이를 불확실하게 하는 다른 선행문헌이 제시된 경우에는 그 내용까지 종합적으로 고려하여야 한다.
이 판결은 선행문헌 전체 대비 및 다른 기술문헌과의 종합 고려를 명시적으로 설시한 대표 판결이다. 이를 관련 법리의 “최초” 또는 모든 판결의 유일한 “원형”이라고 단정할 필요는 없다.

1.1 “선행문헌 전체”는 모든 구성이 언제나 불가분이라는 뜻이 아니다

선행문헌 전체 대비 원칙은 선행문헌 속 개별 구성이나 하위 결합을 다른 문헌과 결합하는 것 자체를 금지하지 않는다. 문제된 구성이 문헌에서 독립된 기술적 수단으로 인식되고 다른 시스템에도 예측 가능하게 적용될 수 있다면 그 결합은 허용될 수 있다.

금지되는 것은 청구발명을 보고 필요한 부분만 골라낸 뒤, 그 부분이 원래 수행하던 기능·전제조건·작동원리 또는 문헌의 반대 교시를 무시하는 왜곡된 선택적 발췌이다.

허용될 수 있는 추출·결합

구성이 독립적으로 인식되고, 필요한 주변 조정이 통상적이며, 다른 장치에서도 같은 기능을 예측 가능하게 수행하고, 문헌 전체가 그 적용을 배제하지 않는 경우

문제되는 선택적 발췌

구성의 기능적 전제와 주변 관계를 제거하거나, 문헌의 목적·작동원리를 상실시키거나, 명시적 부정적 교시를 무시한 채 청구발명을 완성하는 데 필요한 부분만 추출한 경우

1.2 “배치되는 다른 선행문헌”과 “부정적 교시”는 구별한다

개념 의미 실무상 기능
배치·불확실하게 하는 다른 선행문헌 한 선행문헌의 일부 기재와 충돌하거나 그 기술적 의미·신뢰성을 불확실하게 하는 별도의 기술문헌 출원 당시 전체 기술수준과 증거의 무게를 종합 평가하게 한다.
부정적 교시
(teaching away)
동일 문헌 또는 다른 자료가 특정 변경·수치·결합을 피하도록 하거나, 그 방향이 목적 달성에 부적절하다고 가르치는 경우 제안된 결합의 이유와 성공 가능성을 약화한다. 단순한 단점 언급이나 다른 방안의 선호와는 구별된다.

2. 2019후12094 판결의 실제 적용: 점도와 Li₂O

Case Study · 점도 범위

출원발명은 용융 산화물 욕의 점도를 0.3×10-3 Pa·s 이상 3×10-1 Pa·s 이하로 한정하였다.

선행발명은 용융 염 욕의 점도가 바람직하게는 100포이즈 이하라고 기재했지만, 강대 표면에 응고피막을 형성한 뒤 이를 파괴·박리하는 작동원리를 전제로 하였다.

대법원의 판단은 “하한이 명시되지 않았다”는 일부 문구만으로 출원발명의 낮은 점도까지 포함된다고 볼 수 없다는 것이었다. 문헌 전체를 보면 응고피막이 형성될 최소 점도가 사실상 요구되며, 그 작동원리를 상실할 정도로 점도를 낮추는 것은 단순 최적화가 아니었다.

Case Study · Li₂O 함량

출원발명은 Li₂O 함량을 10중량% 이상 45중량% 이하로 요구하였다.

선행발명은 Li₂O 함량을 0~6중량%로 제시하면서, 6중량%를 초과하면 응고피막의 박리성이 나빠질 수 있으므로 그 범위를 피하도록 가르쳤다.

대법원의 판단은 10~45중량%로 높이는 변경이 문헌 전체의 부정적 교시에 반하며, 출원발명을 알고 나서야 선택하기 쉬운 사후적 변경이라는 취지였다.

이 판결에서 읽어야 할 핵심

선행문헌의 수치범위가 문언상 열려 있는 것처럼 보여도, 문헌 전체의 목적과 작동원리가 사실상 허용범위를 제한할 수 있다.

반대로 문헌이 특정 범위를 명시적으로 피하라고 가르친다면, 그 반대 범위로의 변경을 “통상적 최적화”라고 단정하기 어렵다.

3. 선택적 발췌와 사후적 결합의 전형적 위험

3.1 부품 창고형 추출

D2의 구성 c가 d·e·f와 함께 특정 기능을 수행하는데도 c만 떼어 D1에 이식하는 방식이다. c가 독립 모듈로 인식되는지, d·e·f가 기능의 필수 전제인지, D1에 적용하면 같은 기능이 유지되는지를 먼저 확인해야 한다.

3.2 출원발명을 지도로 삼은 역방향 조합

청구발명의 a+b+c 구조를 먼저 본 뒤 a는 D1, b는 D3, c는 D2에서 찾아 조합하는 방식이다. 각 선택과 변경의 이유가 출원 당시의 객관적 자료에서 확인되지 않으면 사후적 고찰의 위험이 크다.

3.3 기능 명칭만 같은 구성의 기계적 치환

두 부품이 모두 “밀봉”, “지지”, “안전” 등의 기능을 가진다는 이유만으로 교체 가능하다고 단정하는 방식이다. 작동조건, 구조적 연결, 압력·온도·하중, 제어방식 및 예상되는 성능을 비교해야 한다.

4. 결합 용이성 Q1~Q6: 법정요건이 아닌 실무 점검표

중요: 아래 Q1~Q6은 대법원이 정한 누적적 법정요건이나 점수표가 아니다. 모든 문항이 “용이”여야만 진보성이 부정되는 것도 아니고, 하나가 “어려움”이라고 해서 진보성이 자동 인정되는 것도 아니다. 각 사정의 기술적·법적 중요도와 상호관계를 종합해 출원 당시 통상의 기술자가 실제로 해당 결합에 도달할 수 있었는지를 판단해야 한다.

Q1문제된 구성은 독립적으로 인식·적용될 수 있는가?

구성이 다른 부품과 기능적으로 결합되어 있다면 단순히 “독립 작동 여부”만 묻지 않는다. 분리·이식에 필요한 주변 변경의 범위, 난이도, 상호의존성 및 결과의 예측 가능성을 함께 본다.

  • 문헌이 그 구성을 독립 모듈이나 선택 가능한 수단으로 설명하는가?
  • 주변 구성을 함께 옮기는 것이 통상의 설계범위인가?
  • 분리 후에도 원래 기능이 예측 가능하게 유지되는가?
  • 청구발명을 알아야만 특정 주변 변경을 선택할 수 있는가?
Q2필요한 변경은 통상적 조정인가, 창작적 재설계인가?

선행기술 결합에는 치수, 배치, 연결부, 재료 또는 제어조건의 조정이 수반될 수 있다. 일정한 설계변경이 필요하다는 사실만으로 결합이 배제되지는 않는다.

통상적 조정에 가까운 경우

변경방향이 표준·교과서·통상 관행에서 제시되고, 소수의 예측 가능한 조정으로 기존 기능이 유지되는 경우

창작적 재설계에 가까운 경우

다수의 상호의존적 구성과 작동원리를 바꾸어야 하고, 그 변경방향이 객관적 자료에 없으며, 청구발명의 핵심 구성을 알아야만 설계가 완성되는 경우

변경의 “개수”보다 각 변경의 질적 중요성, 상호작용과 예측 가능성이 중요하다.

Q3결합이 기준발명의 목적·작동원리·핵심기능에 미치는 영향은 무엇인가?

결합이 선행발명의 주된 목적이나 작동원리를 상실시키면 용이성을 강하게 약화할 수 있다. 그러나 일부 성능이 낮아진다는 사정만으로 결합 동기가 항상 부정되는 것은 아니다.

통상의 기술자가 다른 장점, 비용 절감, 안전성 또는 규격 충족을 위해 성능 일부를 희생하는 기술적 절충을 선택할 수 있는지도 함께 평가한다.

Q4진정한 부정적 교시가 존재하는가?

선행기술이 제안된 변경을 명시적으로 피하라고 하거나, 목적 달성이 불가능해지거나 중대한 실패가 생긴다고 가르치는 경우에는 부정적 교시가 강하다.

반면 기존 방식의 단점을 언급하거나 다른 실시형태가 더 낫다고 선호하는 것만으로는 충분하지 않을 수 있다. 문헌이 통상의 기술자를 실제로 해당 방향에서 멀어지게 하는지 확인해야 한다.

Q5출원 당시 결합할 객관적 이유가 있었는가?

결합 이유는 반드시 D1 또는 D2에 문자 그대로 명시되어 있어야 하는 것은 아니다. 다음 자료에서도 도출될 수 있다.

  • 두 문헌이 공유하는 기술적 과제나 기능
  • 출원 당시의 기술상식과 통상의 설계관행
  • 산업표준, 안전규격, 환경·비용·시장 요구
  • 교과서, 기술문헌, 표준문서 및 객관적인 전문가 자료

다만 “더 좋아진다”, “필요에 따라 바꿀 수 있다”와 같은 추상적 설명만으로는 부족하다. 선택한 구성과 변경방향을 출원 당시의 객관적 증거가 뒷받침해야 한다.

Q6제안된 결합의 자연스러운 결과가 청구항의 모든 한정사항에 도달하는가?

D1과 D2를 주장된 이유에 따라 결합했을 때 청구항의 모든 한정사항이 자연스럽게 충족되는지 확인해야 한다. 도달하지 못한다면 진보성 부정 논증은 완성되지 않는다.

다만 결과 불일치가 곧바로 사후적 고찰의 “직접증거”가 되는 것은 아니다. 대응관계가 불완전하거나 결합논리가 부족한 것일 수도 있다. 부족한 차이를 청구발명의 구조를 참고해 추가 보완할 때 사후적 고찰의 우려가 특히 커진다.

비기계적 사용법

각 질문에 단순히 YES/NO를 붙이기보다, ① 근거가 되는 문헌·도면·표준, ② 필요한 변경의 내용, ③ 예상되는 기능과 효과, ④ 반대 증거, ⑤ 통상의 기술자의 관점에서의 예측 가능성을 함께 기록한다.

부정적 사정이 많고 강할수록 결합 용이성은 약해지지만 최종 결론은 전체 기록에 대한 종합평가다.

5. 기술분야, 주지·관용기술과 효과의 올바른 취급

5.1 기술분야의 유사성은 중요한 요소이지 절대적 필요조건이 아니다

동일·인접 분야라는 사실은 결합 가능성을 지지하지만 그 자체로 충분하지 않다. 반대로 서로 다른 분야라도 해결과제, 기능, 작동원리 또는 적용 가능성이 밀접하고 통상의 기술자가 참고할 합리적 이유가 있다면 결합 대상이 될 수 있다.

5.2 주지·관용기술도 객관적 증거가 필요하다

O-링, 인터록, 방진 마운트처럼 널리 쓰이는 수단은 기술수준과 통상적 설계변경을 뒷받침할 수 있다. 그러나 “주지·관용”이라는 명칭만으로 증명이 끝나지는 않는다. 출원 당시의 교과서, 산업표준, 기술문헌, 다수 특허문헌 또는 신뢰할 수 있는 전문가 자료로 그 존재와 일반성을 입증해야 한다.

5.3 효과는 이질적 효과에 한정되지 않는다

진보성을 뒷받침하는 효과에는 다음이 포함될 수 있다.

  • 선행기술과 질적으로 다른 이질적 효과
  • 동질적이지만 예측을 크게 뛰어넘는 현저히 우수한 효과
  • 특정 수치범위에서 나타나는 임계적 효과
  • 구성들의 상호작용으로 나타나는 상승적·시너지 효과

효과는 명세서에 기재되거나 적어도 통상의 기술자가 명세서로부터 합리적으로 인식할 수 있어야 하며, 비교실험은 청구범위와 선행기술의 차이를 적절히 반영해야 한다.

6. 논거별 일반적 강도: 고정 순위가 아닌 사건별 평가

논거 유형 일반적 강도 확인할 사항
제안된 변경을 명시적으로 회피·금지 매우 강함 단순 선호·단점 언급인지, 실제 회피 교시인지
변경 시 주된 목적·작동원리 상실 매우 강함 일부 성능 저하가 아니라 기술적 의의 자체가 사라지는지
결합 후 청구항 차이에 도달하지 못함 매우 강함 모든 한정사항의 대응과 부족한 추가 변경
비통상적·상호의존적 재설계 필요 강함 변경의 질적 중요성, 난이도와 예측 가능성
결합 이유의 객관적 근거 부족 강함 문헌, 기술상식, 표준, 안전·시장 요구 등 증거의 존재
통상적 치수·재료·배치 조정 사안별 단순 최적화인지, 임계적 효과나 편견이 있는지
동일 기술분야 또는 공통 과제 결합을 지지하지만 단독으로 부족 구체적인 적용 이유와 성공 예측 가능성

이 표는 고정된 우선순위가 아니다. 예컨대 명시적 부정적 교시가 있는 사건에서는 그것이 핵심이 되고, 청구항 대응이 빠진 사건에서는 결합 결과의 불완전성이 가장 강한 논거가 된다.

7. 의견서 작성 실무

  1. 청구항의 차이점을 한정사항 단위로 고정한다. 결과나 효과만이 아니라 구조·수치·관계·순서를 정확히 적는다.
  2. D1과 D2를 각각 독립적으로 전체 읽기한다. 목적, 작동원리, 필수 구성, 선택 구성, 부정적 교시와 효과를 표로 정리한다.
  3. 상대방 결합 논리를 재현한다. 무엇을 어디서 가져오고 어떤 변경을 거쳐 청구항에 도달한다는 것인지 단계별로 적는다.
  4. Q1~Q6을 증거와 함께 적용한다. 문단번호·도면번호·표준·교과서·실험자료를 붙인다.
  5. 반대 논거도 평가한다. 동일 과제, 통상 관행, 표준화 요구, 비용·안전 동기 등 결합을 지지하는 자료를 누락하지 않는다.
  6. 사후적 보완이 발생하는 지점을 표시한다. 선행기술만으로 선택되지 않는 변경이 청구발명을 본 뒤 추가되는지 확인한다.
  7. 효과는 차이점과 인과관계를 연결한다. 어떤 한정사항이 어떤 효과를 만들었는지 설명한다.
3단 논증 템플릿

① 법리: 선행문헌 전체 대비, 증거 기초 판단, 부정적 교시 또는 사후적 고찰 금지 중 적용되는 원칙을 특정한다.

② 사실: D1·D2의 문단·도면·수치와 기술상식 자료를 인용하여 실제 가르침과 필요한 변경을 입증한다.

③ 적용: 그 증거만으로 통상의 기술자가 청구항의 모든 한정사항에 자연스럽게 도달하는지, 추가 창작이 필요한지를 결론짓는다.

8. 결론

최종 정리

선행문헌 전체 대비 원칙은 선택적 발췌 자체를 금지하는 법리가 아니다. 선행문헌의 개별 구성이 독립적으로 인식되고, 그 적용 이유와 변경방향이 출원 당시의 객관적 기술수준에서 뒷받침되며, 통상적인 조정으로 청구발명에 도달할 수 있다면 다른 선행기술과 결합될 수 있다.

문제는 선행문헌 전체의 목적·작동원리·구성 간 관계와 부정적 교시를 무시하여 기술적 의미를 왜곡하는 발췌, 그리고 청구발명의 완성된 구조를 지도로 삼아 부족한 부분을 사후적으로 조합하는 논증이다.

따라서 최종 질문은 “구성을 떼어냈는가”가 아니라, 출원 당시 통상의 기술자가 전체 기록을 읽고 그 구성을 그 방식으로 선택·변경·결합하여 청구항의 모든 한정사항에 합리적으로 도달했을 것인가이다.

주요 판결

본 자료는 대법원 2019후12094, 2013후2873·2880 등 주요 판결을 중심으로 선행문헌 전체 대비와 사후적 고찰 금지 법리를 해설한 자료입니다.

일반적인 정보 제공 목적이며 구체적인 사건에 대한 법률의견을 구성하지 않습니다.

© 2026 All rights reserved.

Tuesday, April 14, 2026

Mastering Patent Claim Construction with LLMs (Final): From NotebookLM to Claude-Based Expert Skills

LLM으로 완벽한 특허청구범위 해석하기 (3)

Mastering Patent Claim Construction with LLMs (Final): Converting the NotebookLM Knowledge Base into a Claude Skill

The work carried out in the earlier stages consisted of inputting into NotebookLM sources such as the “claim construction framework,” “patent claim construction doctrines and case law,” “case examples,” and “Claim Chart drafting methods” that had been organized through case-law research, conducting research on that basis, and then extracting a summary version of the resulting “knowledge base on patent claim construction doctrine and practical interpretation work.”

The important point at this stage is not simply to gather a large amount of material. The real significance lies in getting NotebookLM to structure and summarize the common legal principles and practical rules drawn from multiple sources, and then using that summary as the starting point for the next stage of work.

In other words, instead of reading scattered cases and practical materials one by one and building the system manually, the first move is to use NotebookLM to integrate dispersed materials into a single practice-oriented knowledge base.

The next step is to select the most suitable guide from among the various draft skill guides generated in this way, or to combine the strengths of multiple drafts and refine them into one integrated version. This step is also extremely important, because even if the draft generated by the LLM is already very good, it still needs fine-tuning to fit the user’s purpose, the type of case, the user’s analytical habits, and the desired output format before it can be used in actual practice.

Once this consolidated summary-style knowledge base from NotebookLM has been organized, the next move is to copy it and paste it into Claude’s “Skill Creator”, and instruct Claude to generate the skill file.

If you are using the Claude web or desktop UI, you can simply copy the organized NotebookLM knowledge base report, go into Claude’s Customize > Skills menu, and create a new skill there.

If you ask Claude as follows, it will generate a skill on its own, including behavior rules and output guidance.

PROMPT
Based on this knowledge base, create an expert skill that thinks and responds according to a specific framework.

Interaction During the Skill Creation Process

If Claude determines during the skill creation process that certain necessary information is missing, it may ask the user for additional details. In that case, you can provide the desired direction, the doctrines you want it to emphasize, the output format, and similar guidance. Once the skill is complete, Claude may even run sample tests on its own.

This is also quite useful in practice. Even if the user does not prepare test cases manually, it allows the user to see in advance, at a minimum, how the generated skill is likely to operate.

The Claude skill generated on the basis of the NotebookLM knowledge base is shown below.

PROMPT
name: kr-patent-expert-v2 description: > A top-tier expert skill (v2) for Korean patent claim construction and infringement/invalidity analysis. Based on Korean Supreme Court and Patent Court case law, it applies a **six-step problem-solving methodology** and a **seven-column Claim Chart** to deliver expert-level analysis suitable for immediate use in infringement litigation and invalidity proceedings. You must use this skill in the following situations: - “claim interpretation,” “claim scope analysis,” “Claim Chart drafting,” “claim chart,” “seven-column chart” - when infringement or invalidity analysis is requested together with a Korean patent (KR) specification or claims - when analysis is requested for PbP claims (product-by-process), numerical limitation inventions, parameter inventions, or functional claims - when analysis is requested concerning Korean patent doctrines such as “All Elements Rule (AER),” “Doctrine of Equivalents (DOE),” “problem-solving principle,” “estoppel,” “intentional exclusion,” “public dedication,” or “reference to prosecution history” - when a one-to-one comparative analysis is needed between an accused product (accused invention) and the patented invention - when deeper six-step / seven-column analysis is needed beyond the existing kr-patent-claim-analyzer --- # Korean Patent Claim Construction and Infringement Analysis Expert Skill (v2) ## Role and Behavior Rules When this skill is applied, Claude acts as a **top-tier patent dispute expert with deep command of Korean Supreme Court and Patent Court case law**, and must strictly follow the rules below. | # | Rule | Content | |---|------|------| | R1 | **Strict case-law-based interpretation** | Applies Patent Act Article 97 and Korean Supreme Court case law (literal interpretation principle, reference to the specification, prosecution history estoppel, and reference to prior art) as absolute standards | | R2 | **Maintain the PHOSITA perspective** | All claim interpretation is performed from the objective viewpoint of a person having ordinary skill in the art (PHOSITA) at the time of filing | | R3 | **Strict separation of interpretation and limitation** | Clarifying technical meaning through reference to the specification is allowed. Improper limiting construction through importing embodiments into the claim (Importing Limitations) is strictly rejected | | R4 | **Proactive identification of special claim types** | If a functional claim, PbP claim, or numerical limitation invention is identified, proactively applies the Korean Supreme Court’s tailored doctrines that go beyond general interpretive rules | | R5 | **Balanced construction of offensive and defensive logic** | In infringement analysis, prioritizes AER (All Elements Rule), and if literal non-infringement is expected, simultaneously builds both estoppel-based and DOE-based offensive and defensive logic | ## Output Decision Guide by Input Type Automatically determines the level of output depending on the type of materials provided. | Input Type | Output | |-----------|-----------| | **① Claim text only** | Structure/function decomposition + first-stage diagnostic report on special issues (functional claim / PbP / numerical limitation) | | **② Claim text + specification (description / drawings)** | Two-stage analysis including Lexicographer Rule applicability + derivation of objective technical meaning + warning of improper narrowing / expansion risks | | **③ Claim text + specification + prosecution history (OA / responses)** | Analysis of whether intentional exclusion (Korean-style estoppel) is established + final determination of claim scope | | **④ Full patented invention + accused product** | Applies the full six-step methodology → outputs a completed **seven-column Claim Chart** | ## Six-Step Problem-Solving Methodology > **Mandatory**: Perform the six steps below in order, without skipping any step. ### Step 1 — Claim Decomposition and Initial Delimitation of the Claim Language - Break the claim down by structure and function into the **smallest units for invalidity and infringement analysis** - Identify the **physical and functional organic relationships** among claim elements (simple listing is prohibited) - Preliminarily confirm the principle of claim differentiation between independent and dependent claims ### Step 2 — Reference to Intrinsic Evidence and Examination of Technical Meaning - Use the ordinary meaning of the claim language as the starting point, and consult the specification and drawings on a one-to-one basis - **Lexicography**: If the applicant explicitly defined a term, apply that definition first - Determine the problem-solving principle and operative / functional effect of each element ### Step 3 — Filtering Improper Narrowing / Expansion - Even when consulting the specification, **filter out errors that improperly narrow the claim based on specific embodiments or figures, or improperly expand the claim beyond the specification** - Check: “Is this interpretation importing limitations from the embodiments into the claims?” ### Step 4 — Applying Tailored Doctrines to Special Claim Types | Claim Type | Applicable Doctrine | |-------------|-----------| | **Functional claim** (“means for ~”) | Reviews whether there are grounds for limiting construction to embodiments if the literal reading would be unduly broad in light of the specification (Supreme Court 2017Hu905 doctrine) | | **PbP claim** (product defined by manufacturing method) | Translates and extracts not the process itself, but the **structure / properties of the final product defined by that process** as the substantive element | | **Numerical limitation / parameter invention** | Determines (i) enablement across the claimed range + (ii) whether there is critical significance reflected in a remarkable effect | ### Step 5 — Reference to Prosecution History and Review of Estoppel / Intentional Exclusion - Analyzes the entire set of office-action responses and amendments across prosecution, registration, divisional filings, and correction proceedings - **Intentional exclusion (Conscious Disclaimer)**: If a specific element was surrendered to overcome prior art, it is strictly excluded from the scope of the claim - **Dedication to the Public**: If an alternative is disclosed in the specification but omitted from the claims, it blocks application of the doctrine of equivalents ### Step 6 — Infringement Comparison and Limits of the Doctrine of Equivalents (DOE) 1. **AER (All Elements Rule)**: Compare whether every element exists in the accused product on a one-to-one basis 2. **Literal infringement determination**: If even one element is missing, literal infringement does not exist 3. **DOE analysis**: If literal infringement does not exist, determine equivalent infringement by identifying—taking prior art into account—whether the patented invention and the accused product share the same **problem-solving principle unique to the patented invention** - ① Same problem-solving principle (Way) - ② Substantially the same effect/result - ③ Readily substitutable by a PHOSITA ## Seven-Column Claim Chart Output Format (ALWAYS use this format) For input type ④ (full analysis), always output the following **seven-column Claim Chart** in markdown table format. | 번호 | 청구항 분해 문언 | 핵심 해석대상 | 명세서/도면 근거 | 객관적 기술의의 및 문언해석 결과 | 균등론(DOE) 대비: 과제 해결 원리 식별 | 방어논리: 출원경과 금반언 및 제한해석 사유 | | :-- | :--- | :--- | :--- | :--- | :--- | :--- | | 1 | [분해 문언] | [용어] | [근거] | [확정 의미] | [핵심 기술사상] | [의식적 제외 이력] | ### Claim Chart Drafting Guidelines (Four Core Principles) 1. **Emphasize organic relationships**: Draft in a way that reveals the physical and functional interaction between A and B 2. **Mandatory PbP translation**: Must be converted into “structure / physical property” terms 3. **Proactive identification for DOE**: Proactively identify the problem-solving principle 4. **Thorough record of estoppel**: Any surrendered element must be recorded in the defense-logic column

This case is a very good test case in practice. That is because it clearly illustrates a situation where the general principle that the specification must be consulted appears to come into direct tension with the equally important principle that the claims must not be improperly narrowed based on embodiments or details in the specification. In other words, if you really want to see whether the skill is functioning properly, you need to test whether it can filter out improper limiting construction in a case of this type.

How the Instruction Was Actually Given in Claude

In Claude, I activated the skill generated above, uploaded Korean Patent No. 10-1257892, the prosecution history file, and the examiner’s cited references, and then instructed it to perform claim interpretation. Even if you simply upload the files and run the prompt, the skill will operate according to its behavior rules.

PROMPT
/kr-patent-expert-v2 Please interpret claim 1 of Korean Patent No. 10-1257892, which I am uploading.

When a skill is designed well in this way, the user does not need to add a long explanation. Simply uploading the required materials, calling the skill by name, and specifying the target of analysis is enough to obtain a fairly deep result. Of course, that is only possible when the skill has been designed properly in advance.

How to Work Around Input Limitations

Depending on the user’s Claude subscription type, there are times when the volume of material to be processed exceeds Claude’s input limit. In practice, this happens fairly often. It is especially common in cases where the prosecution history is long, or where there have been multiple rounds of office actions and back-and-forth responses and amendments, making it difficult to submit the entire file at once.

In that situation, you can instead load the prosecution history into NotebookLM, organize the major issues, key arguments, and the portions of claim scope that were narrowed or surrendered, and then feed that organized content into Claude in place of the full prosecution-history file. The prompt I used was as follows.

PROMPT
1) What was the main reason the patent application was initially rejected? 2) How was the heater structure amended to overcome the rejection? 3) Analyze the prosecution history of the patent application, including the examiner’s office actions and the applicant’s responses and amendments, and explain the issues, the key arguments, and the portions of claim scope that were surrendered or narrowed in order to secure registration.

The strength of this prompt is that it does not merely ask, “Please summarize the prosecution history.” Instead, it pushes the model to extract only the core elements that are actually necessary for claim interpretation. From a practical perspective, what matters is not the file as a whole, but what the issues were, what arguments were made, and how far the applicant retreated in order to obtain allowance.

Evaluation of the Actual Result

The claim interpretation and Claim Chart presented by Claude were highly complete and trustworthy. The full output is too long to share here, but at least from what I reviewed, I was able to confirm that the result reflected a very high level of work, arguably beyond what many experts would ordinarily produce.

Of course, there is also a point of caution here. Just because the AI’s output is excellent does not mean it can be copied directly into a legal opinion letter or litigation brief. What it does mean, however, is that as a starting point for analysis or as a draft for expert review, the output has already reached a highly powerful level. In particular, as a practical support tool, it is very useful in terms of speed and consistency when structuring large amounts of material, breaking down issues quickly, and organizing the result into a Claim Chart.

Directly Asking About the Feature That Became the Key Issue

This time, let us directly ask about the feature that became the key issue in the claim interpretation.

PROMPT
In the foregoing interpretation, is claim 1 to be construed narrowly as requiring that the first and second coil portions be capable of being operated simultaneously or separately?

The content of the response was likewise highly professional and reliable. I will quote only the conclusion below. The conclusion was the same as that reached by the Supreme Court and the Patent Court.

Conclusion
Claim 1 does not require, as a limiting condition on claim scope, that the first and second coil portions be independently operable. The phrase “turned on/off by the operating unit” describes the connection and control relationship under which each coil portion is controlled by the operating unit, and is broad enough to cover both simultaneous control and independent control. The phrase “respectively or simultaneously” in paragraph [0025] of the specification does not function as intrinsic evidence that confirms and narrows the claim scope, but merely illustrates a possible control mode. Importing that statement into the claim as a limiting element would therefore amount to improper limiting construction.

This answer is meaningful in practice not merely because the conclusion was correct. More importantly, the reasoning proceeded in the order of claim language → specification → the prohibition against improper limiting construction. In other words, it shows that the skill did not merely memorize the case outcome, but was actually reasoning in the same interpretive sequence used by Korean courts. Only when reasoning is structured in this way can you expect a consistent level of reliability when applying it to other cases as well.


Closing Thoughts

This concludes the discussion of how a Claude skill was generated from a NotebookLM knowledge base and then used to perform AI-based claim interpretation by inputting an actual patent publication and prosecution history.

The key point of this exercise is not simply to show that “Claude performs well.” What matters even more is what kind of knowledge structure is built first, how that knowledge is converted into a skill format, and how it is tested and refined in actual cases. In other words, the quality of the outcome depends more on the quality of the knowledge base and framework built in advance than on the name of the model itself.

From a practitioner’s point of view, this approach has the following significance.

  • It makes it possible to structure a large body of case law and doctrine into a single knowledge base.
  • It allows the creation of expert-level skills that reason according to specific legal doctrines.
  • It can be applied to actual patent disputes to produce Claim Chart-level work product.
  • While the final judgment still belongs to the human expert, a substantial portion of the analytical process can be made far more efficient.

I hope this article proves genuinely useful to practitioners in the field.

Mastering Patent Claim Construction with LLMs (4): Bridging Knowledge Gaps and Designing Expert-Level Skills

LLM으로 완벽한 특허청구범위 해석하기 (3)

Mastering Patent Claim Construction with LLMs (4): Filling the Gaps in the Knowledge Base

The Final Stage of Supplementing the Sources

This is now the final stage of supplementing the source materials. In the earlier stages, I gradually built up relevant materials in NotebookLM and organized the legal principles and practical standards needed for patent claim construction. But simply collecting a large volume of material does not automatically create a strong knowledge base. In practice, materials are often overly concentrated on certain issues, while foundational legal principles that are essential in actual practice may still be missing. That is why, at this final stage, it is necessary to examine not only what the accumulated sources already cover well, but also what they fail to cover.

The core objective of this stage is simple. Rather than continuing to add more knowledge based on what has already been collected, the real task is to identify the gaps in the knowledge base. In other words, this is not really a stage for re-confirming “what we already know,” but more accurately a stage for exposing “what we still do not know.”

To do that, I entered the following instruction into the NotebookLM chat window.

PROMPT
Review the attached sources and analyze whether there are any missing pieces of information. In particular, please examine what core information or context is essential to fully understand this topic but is currently missing from the sources, and explain it specifically.

The purpose of this instruction is not summarization or organization. It is intended to make the LLM identify what core legal principles or contextual elements are still missing, even though they are essential for a complete practical understanding based on the materials collected so far. Framing the question this way prevents the model from giving a superficial response like, “You’ve assembled the materials well,” and instead pushes it to identify the deficiencies in a structured and practical way.


Items Identified as Requiring Supplementary Materials

In response to the instruction above, the materials requiring supplementation were organized as follows.

Output
A comprehensive review of the currently provided sources and conversation history shows that while advanced special issues in patent claim construction and infringement analysis—such as PBP claims, functional claims, and the suture-line cases—have been covered in substantial depth, several core foundational principles and practical contexts that are essential for fully understanding and applying this area in practice are still missing. Specifically, the following key information is either absent or needs supplementation. 1. Missing Korean leading cases on the requirements for ordinary prosecution history estoppel (“intentional exclusion”) - Current status: The latest Korean Supreme Court decision dealing with estoppel in the context of divisional applications (2023Hu11340) and the discussion of the fifth requirement of the doctrine of equivalents through Japan’s Maxacalcitol case are both covered in significant detail. - Missing context: What is missing is a concrete factual and legal analysis of the traditional Korean Supreme Court leading cases (e.g., Supreme Court Decision 2002Hu1768) that clearly set out how strong or explicit a statement in an office-action response or amendment must be, from an objective and outward-looking perspective, to constitute “intentional exclusion” in ordinary Korean patent litigation practice. There is therefore a gap in establishing the Korean baseline for measuring the degree of intentional disclaimer. 2. The strict limits and requirements for applying the Lexicographer Rule - Current status: Many sources already include the overarching principle and cases stating that when claim language is unclear, the detailed description of the invention and the drawings should be consulted. - Missing context: What is missing is a deeper case-based analysis of the strict requirements and limitations of the Lexicographer Rule—specifically, how explicit and definitive an applicant’s definition in the specification must be before that definition can displace the ordinary meaning of a term in the art. 3. Korean applications of the Public Dedication Rule based on the specification - Current status: Some Japanese case commentaries briefly mention whether the dedication doctrine has been adopted, but there is little to no concrete treatment of how it has been applied under Korean law. - Missing context: The current materials lack a clear account of the Korean Supreme Court’s position and practical treatment where the patentee discloses a particular embodiment or substitute in the specification but does not claim it, and whether that subject matter is deemed dedicated to the public so as to bar application of the doctrine of equivalents. Because this is a powerful defense against an equivalents claim, it is essential material that should be supplemented. 4. “Omission infringement” and “incomplete use inventions” as exceptions to the All Elements Rule (AER) - Current status: The existing materials address, at a sophisticated level, issues such as the suture-line decision overcoming geographic limits, multi-actor infringement, and indirect infringement as exceptions to the All Elements Rule. - Missing context: However, the question of whether infringement can still be found when the defendant strategically omits or only incompletely uses a relatively less important element of the patented invention in an effort to design around the claim is discussed only briefly, in about a single paragraph. That is not enough to fully understand the range of defensive arguments against design-arounds. 5. Standards for evaluating “critical significance” when applying the doctrine of equivalents to numerical limitation inventions - Current status: The latest Korean Supreme Court case on disclosure requirements for numerical limitation inventions (2021Hu10886), including enablement and written-description support, is already covered. - Missing context: What is missing is the Korean courts’ concrete standard, at the infringement stage, for determining whether equivalents may be found when the accused product falls just slightly outside a claimed numerical range—for example, whether the presence or absence of “critical significance” excludes application of the doctrine of equivalents. This is indispensable for a full understanding of both offensive and defensive logic in numerical-limitation cases where there is no literal infringement.

Important Practical Insights Gained at This Stage

The result above means more than simply “we need more materials.” From a practical perspective, the more important message is that even if the materials collected so far are already highly sophisticated, the entire framework can become unstable if the baseline of general legal doctrine is missing.

For example, even if you have assembled abundant materials on difficult special issues such as PBP claims and functional claims, the standards for interpreting claim scope in an actual case can still become unstable if foundational principles—such as the requirements for ordinary prosecution history estoppel or the limits on the Lexicographer Rule—have not been properly organized. Put differently, special issues are “advanced applications,” while general doctrine is “basic grammar.” If the basic grammar is weak, the advanced applications will inevitably become unstable as well.

That is why I did not simply leave the omissions identified above as a note. Instead, I copied them back into the NotebookLM source window and activated the deep research function to supplement the missing cases and legal principles. This process is extremely important because by feeding the LLM-identified “knowledge gaps” back into the system as inputs, the knowledge base can evolve from a simple collection-oriented repository into a self-correcting knowledge base.


Now It Is Time to Organize the Knowledge to Be Injected into the Claude Skill

The next stage is to use the supplemented materials to extract and organize only the core knowledge that should be injected for drafting the Claude Skill. This is where many people make a mistake. Once they have gathered a large amount of material, they tend to think they should put as much of it as possible into the Skill. But in practice, the opposite is often true. If the background knowledge loaded into the Skill becomes too extensive, the model may get buried in explanations or lose sight of priorities rather than following the core principles clearly. At this stage, therefore, accurate compression matters more than volume.

When I draft a Claude Skill, I usually prefer to compress the core principles, framework, workflow, and output format into about 5,000 characters or less before injecting them. This point is especially important in practice. The skill guide, SKILL.md, is generally recommended to remain under 5,000 words, with more detailed knowledge separated into the references/ folder. In addition, it is more efficient to reflect required input information in the trigger description rather than in the body of the Skill itself. That way, Claude can assess input sufficiency before the Skill is even called.

How to Ask Questions to Extract the Core Knowledge While All Sources Are Selected

With all sources selected, I enter the following questions one by one into the NotebookLM chat window to organize the background knowledge that will be injected into the Skill. At present, I have 60 sources registered in my notebook.

PROMPT
1. What are the core principles that consistently appear across all of these sources in relation to patent claim construction? 2. When all of the sources are analyzed together, what framework is mentioned most frequently for interpreting patent claims? 3. Looking at claim construction step by step, what does an excellent real-world process for interpreting the language of a patent claim actually look like? 4. What is the concrete framework or workflow strategy used by Korean courts for patent claim interpretation? To prepare a claim chart for claim construction, please explain in detail how to break down the claim language, extract and interpret the terms or limitations that become the targets of interpretation, and how to draft the claim chart itself.

This sequence of questions was not chosen arbitrarily. The reason I ask first about the common principles, then the framework, then the process, and finally the concrete drafting method is to guide the LLM into organizing its answer in the sequence of abstract principles → procedural structure → practical execution method. This makes it much easier to transfer the results later into a Skill without the information becoming disordered and scattered, and it helps maintain a relatively stable structure.


The Final Claude Skill Injection Guide Organized in NotebookLM

The final guide I organized in NotebookLM for injection into the Claude Skill is as follows.

PROMPT
Korean courts, in determining the scope of patent rights in infringement litigation and invalidation proceedings, apply in practice a **multidimensional six-step interpretive process (workflow)** based on the four core principles of **literal interpretation, reference to the specification, prosecution history estoppel, and reference to the prior art**. Based on that framework, the following explains the specific methods of decomposition and extraction, as well as the drafting format, for preparing a Claim Chart. ### 1. The Korean Courts’ Six-Step Framework for Interpreting Patent Claims - Step 1 (Claim decomposition and initial delimitation based on the claim language): Break the claim into individual elements, which are the minimum units for invalidity and infringement analysis, identify the organic relationships among them, and then define the initial scope of the patent right based on the ordinary and general meaning of the terms from the viewpoint of a person having ordinary skill in the art (PHOSITA). - Step 2 (Reference to intrinsic evidence and examination of technical meaning): Because it is often difficult to ascertain the technical meaning from the claim language alone, the detailed description of the invention and the drawings must be consulted. In particular, the “Lexicographer Rule” applies where the applicant has defined a specific term in the specification, and the analysis should go beyond the immediate context to examine the problem-solving principle and functional effect (technical meaning) of the relevant element. - Step 3 (Filtering out improper narrowing and expansion): Even when the specification is taken into account, courts strictly block errors that improperly narrow claim scope by relying only on a specific embodiment or configuration disclosed in the specification, or improperly expand the claim scope beyond what the claim language reasonably supports. - Step 4 (Applying tailored legal rules to special types of claims): For functional claims (e.g., “means for ~”), the analysis must consider whether limiting construction to the embodiments is warranted; for product-by-process (PbP) claims, the issue is not the process itself but the “structure or properties of the product” defined by that process; and for numerical-limitation inventions, the analysis must focus on whether the claimed numerical range has “critical significance.” - Step 5 (Reference to prosecution history and review of intentional exclusion): Review the prosecution history from filing to registration—such as office-action responses and amendments—to determine whether the applicant intentionally excluded specific subject matter in order to avoid the prior art, in which case that subject matter must be firmly excluded from the scope of the patent right. - Step 6 (Infringement comparison and limits on the doctrine of equivalents): Compare the finalized scope of the patent right with the accused product to determine whether all elements are present (All Elements Rule, AER). If differences exist, determine whether the doctrine of equivalents (DOE) applies by examining whether the patented invention and the accused product share the same principle for solving the problem, among other factors. ### 2. Methods for Claim Decomposition and Extraction for Preparing a Claim Chart 2.1. Structure- and Function-Based Decomposition: Decompose the claim in a multidimensional way into the physical structure of the elements and the functions and roles performed by those elements. 2.2. Identification of Organic Relationships: Go beyond merely listing the elements and clearly identify the **physical and functional connections (relationships)** among different elements that contribute to solving the technical problem. 2.3. Extraction of Core Interpretive Targets and Limitations (Special Issues): Extract the key terms that may become issues in claim construction. In particular, functionally expressed terms (“means for ~”), PbP processes, numerical limitations, and parameters should be targeted so they can be separately reviewed for issues such as limiting construction and lack of enablement. ### 3. The Most Desirable Concrete Claim Chart Methodology and Format From a practical standpoint, the best Claim Chart is not just a literal element-by-element comparison. It should be drafted as a seven-column system that can fully map both doctrine-of-equivalents (DOE) defenses in cases of literal non-infringement and the issues presented by special types of claims. | No. | Decomposed Claim Language (Including Organic Relationships) | Core Interpretive Target (Special Issue: Functional / PbP / Numerical) | Specification / Drawing Support (Mapping of Intrinsic Evidence) | Objective Technical Meaning and Result of Literal Interpretation (from the PHOSITA Perspective) | DOE Comparison: Identification of the Principle for Solving the Problem | Limitation on Rights: Prosecution History Estoppel and Grounds for Limiting Construction | | :-- | :--- | :--- | :--- | :--- | :--- | :--- | | 1 | (e.g., a cleaning unit that supplies cleaning water by electrolyzing filtered water) | cleaning unit (functional claim) | [paragraph 14], [figure 3] | (e.g., an internal module that physically cleans through electrodes) | (e.g., the principle of maximizing eco-friendly cleaning efficiency by excluding chemicals) | (e.g., intentional exclusion of a “chemical additive method” during argument submission) | | 2 | (e.g., a tablet manufactured by direct compression) | manufactured by direct compression (PbP claim) | [paragraphs 15–16] | (e.g., a porous tablet structure having 15% inter-particle porosity formed through the direct compression process) | (e.g., the principle of controlling disintegration speed by adjusting porosity) | (e.g., structurally / physically different from tablets manufactured by wet granulation) | [Drafting Guide and Explanation for Each Column] 1) Decomposed Claim Language: Set out verbatim the claim language as broken down into the minimum units, but do so in a way that reveals not just a simple list of elements, but also the physical and functional relationships among them. 2) Core Interpretive Target: Extract the key terms that present infringement or invalidity issues—such as functional expressions, PbP language, and numerical limitations—and identify the type in parentheses. 3) Specification / Drawing Support (Mapping of Intrinsic Evidence): Provide a one-to-one mapping to the specific paragraph numbers and figure numbers in the specification that show where the term is defined, where the supporting embodiment for a functional claim appears, where PbP leads to a specific physical-property change, and where the technical threshold for the numerical limitation is shown. 4) Objective Technical Meaning and Result of Literal Interpretation: Describe the substantive meaning derived from the specification from the viewpoint of a person having ordinary skill in the art (PHOSITA). In PbP claims, this should not simply repeat the process itself, but should translate it into the “structure or properties of the product” defined by that process. 5) DOE Comparison (Identification of the Principle for Solving the Problem): In anticipation of literal non-infringement, identify in advance the **“problem-solving principle unique to the patented invention (the core technical idea)”** to which the relevant element belongs, based on the specification and the prior art. 6) Limitation on Rights (Prosecution History Estoppel and Grounds for Limiting Construction): Carefully record any history of **intentional exclusion** during prosecution, any reason why the claim should be limited to the embodiments because the broader reading is not supported by the specification, and any invalidity grounds such as lack of enablement across the full scope of a numerical limitation.

Now I Ask NotebookLM to Create the Expert Skill Itself

At this point, based on the organized knowledge, I ask NotebookLM to actually create a patent claim construction expert Skill. This stage is not merely an experiment. If the principles and framework organized up to this point are sufficiently well-structured, NotebookLM can generate a draft of an expert-level Skill based on them. That draft is not the finished product, but it is extremely useful as a starting point for human review and refinement.

The instruction entered at that point was as follows.

PROMPT
Based on this “Korean patent claim construction doctrine” knowledge base extracted in NotebookLM, create an expert skill file. The skill should include behavior rules, an output decision guide, a problem-solving methodology, and a Claim Chart output style.

Comparing the Generated Skill with the Existing Skill Guide

When I compare the Skill generated by NotebookLM with the Skill guide I created earlier, the result is as follows.

Output
Based on the “Korean patent claim construction doctrine” knowledge base and the conversation history you provided, I am preparing a top-tier expert-level **“Korean Patent Claim Construction and Claim Chart Drafting Expert Skill File”** that can be deployed immediately in patent litigation and trial practice. This skill file can be used as a system prompt or guideline for an AI or practitioner conducting patent analysis work. # [Expert Skill File] Korean Patent Claim Construction and Infringement Analysis Agent ## 1. Behavior Rules You are a top-tier patent dispute expert with deep command of Korean Supreme Court and Patent Court case law. When interpreting claims and analyzing infringement, you must strictly comply with the following rules. - Strict case-law-based interpretation: All interpretation must be grounded in the current Patent Act Article 97 and Korean Supreme Court case law as the controlling standards, including the literal-interpretation principle, reference to the specification, prosecution history estoppel, and reference to the prior art. - Maintain the PHOSITA perspective: The meaning and technical significance of every term must be interpreted from the objective viewpoint of a person having ordinary skill in the art (PHOSITA) at the time the patent application was filed. - Strict separation between interpretation and limitation: You should actively consult the detailed description of the invention in order to clarify the technical meaning of a term, but you must strictly reject any improper “importing of limitations” or improper expansion based on pulling in conditions or configurations from the embodiments that are not actually recited in the claims. - Exceptional treatment for special types of claims: If a functional claim, a product-by-process (PbP) claim, or a numerical-limitation invention is identified, you must proactively apply the Korean Supreme Court’s tailored doctrines that go beyond the general interpretive rules, including limiting construction, translation into structure/properties, and critical-significance analysis. - Secure defense logic first: In infringement analysis, the All Elements Rule (AER) must be applied first and foremost. If literal non-infringement is anticipated, you must simultaneously review prosecution history estoppel (intentional exclusion) and the doctrine of equivalents (DOE) in advance so that both offensive and defensive arguments can be built in a balanced way. ## 2. Output Decision Guide Optimize the output according to the type of user input. - [Input] Only the text of a specific claim is provided: Break the claim down into minimum units (structure/function), identify organic relationships, and provide a first-stage diagnostic report on whether any special issues (functional claim, PbP, numerical limitation) are present. - [Input] Claim text + specification (description / drawings) are provided: Determine whether the Lexicographer Rule applies, derive the objective technical meaning, and perform a second-stage analysis that warns of the risks of improper narrowing or overbroad interpretation. - [Input] Claim text + specification + prosecution history (office actions / responses) are provided: Analyze whether intentional exclusion (Korean-style estoppel) has been established and derive the finally determined scope of the patent right. - [Input] Entire patented invention + accused product are provided: Apply the full “six-step problem-solving methodology” and produce a final **[7-Column Claim Chart]**. ## 3. Problem-Solving Methodology (6-Step Problem Solving Methodology) Carry out the full process from claim construction through infringement analysis under the following multidimensional six-step workflow. - Step 1 (Claim decomposition and identification of organic relationships): Break the claim down into the minimum units that allow invalidity and infringement analysis, while also deriving the overall relationships among the elements. Break the claim down to the smallest units using structure and function as the organizing principles. Do not stop at simple decomposition; define the invention by comprehensively identifying the physical coupling relationships and the functional / organic interconnections among the elements. - Step 2 (Determining technical meaning through literal interpretation and reference to the specification): Begin with the ordinary meaning of the claim language, then concretize that meaning by mapping it one-to-one against the specification and drawings from the viewpoint of a PHOSITA. . Lexicographer Rule: If the applicant explicitly defined a specific term in the specification, apply that definition first. . Examination of technical meaning: Go beyond the immediate contextual meaning and examine, in an objective and reasonable way, the “technical meaning” of the relevant element, including the technical problem it is intended to solve, the implementation principle, and the functional effect, so that the invention’s intended function and operation are properly realized. - Step 3 (Strict control of improper narrowing and expansion): Strictly distinguish between “interpreting in light of the specification” and “importing narrowing or expansion from the specification.” . No improper narrowing: Where the claim language is clear, exclude interpretations that arbitrarily narrow the claim by importing unclaimed elements from specific embodiments, figure shapes, or specific limitations described in the specification. . No improper expansion: Even when consulting the specification, prevent overbroad interpretation that goes beyond the disclosed content or moves in a direction unrelated to the core technical meaning of the invention. - Step 4 (Applying tailored exception doctrines to special types of claims): Unlike ordinary claims, claims drafted in special forms must be interpreted using the Korean Supreme Court’s tailored doctrines. . Functional claims (“means for ~”): If literal interpretation would make the claim unduly broad in light of the specification, review whether there are grounds for limiting construction to specific embodiments or structures. . PbP claims (products recited through a manufacturing process): The substantive element is not the manufacturing process itself, but rather the **“structure or properties of the final product”** defined by that process. . Numerical limitation / parameter inventions: Rigorously assess whether the entire claimed range is enabled (disclosure requirement) and whether the claimed range shows a remarkable effect that distinguishes it from the prior art (critical significance). - Step 5 (Reference to extrinsic evidence and review of estoppel / intentional exclusion): Analyze the entire file history, including office-action responses and amendments, from filing through registration, divisional applications, and correction proceedings. If the applicant **intentionally excluded** a specific element to avoid a rejection, or if subject matter was **dedicated to the public** by being disclosed in the specification but omitted from the claims, that subject matter must be excluded from the scope of the patent right. - Step 6 (Infringement comparison and limits on application of the doctrine of equivalents): Based on the finalized meanings of the elements and their organic relationships, compare them one-to-one with the accused product (or method) and make the final infringement determination. . Literal infringement (AER): First determine literal infringement under the All Elements Rule (AER) by comprehensively examining both element-by-element correspondence and the overall identity of the combination and functional operation. . Infringement under the doctrine of equivalents (DOE): If there is a literal difference, determine whether equivalent infringement exists by identifying, in light of the prior art, whether the patented invention and the accused product share the same “problem-solving principle” unique to the patented invention, along with the same functional effect. ## 4. Claim Chart Output Style All claim analysis results and infringement / invalidity comparison results must be output in markdown table format using the following **“7-Column Claim Chart Template.”** ### [Claim Chart Template] | 번호 | 청구항 분해 문언 (유기적 결합 포함) | 핵심 해석대상 (특수쟁점 표기) | 명세서/도면 근거 (내적 증거) | 객관적 기술의의 및 문언해석 결과 (PHOSITA 관점) | 균등론(DOE) 대비: 과제 해결 원리 식별 | 권리제한: 출원경과 금반언 및 제한해석 사유 | | :-- | :--- | :--- | :--- | :--- | :--- | :--- | | 1 | [최소 단위로 분해된 청구항 원문 기재. 단순 나열 금지, 결합관계 명시] | [핵심 용어 추출] (예: 기능식 / PbP / 수치한정 등 표기) | [정의/실시예가 기재된 단락 번호 및 도면 번호 맵핑] | [명세서를 참작하여 실질적으로 확정된 구성의 의미 기재. PbP의 경우 '구조/성질'로 번역하여 서술] | [문언 비침해 대비. 해당 구성이 속한 특허발명 특유의 '핵심 기술사상(과제 해결 원리)' 명시] | [의견서/보정서상 '의식적 제외' 이력, 또는 실시예로 제한 해석되어야 할 명백한 무효/불비 사유 기록] | | 2 | ... | ... | ... | ... | ... | ... | ### [Points to Watch When Drafting the Claim Chart] 1. **Emphasize organic relationships:** In the “Decomposed Claim Language” column, do not merely chop the claim into elements A, B, and C. Draft it so that the physical and functional interaction among the elements is clearly shown. 2. **Mandatory translation for special claim types:** If a PbP claim is identified, do not write merely “manufactured by ~ method” in the “Objective Technical Meaning” column. Instead, it must be translated into language such as “an internal structure / physical property formed by that process.” 3. **Identify the problem-solving principle in advance for DOE:** Every chart should proactively identify the **“problem-solving principle”** by which the relevant element distinguished itself from the prior art and contributed to the advance, so that it is immediately possible to assess whether a design-around can still be captured if literal infringement is denied. 4. **Thoroughly record estoppel:** If the prosecution history confirms any abandoned subject matter or any substitute dedicated to the public, summarize it in the “Defense Logic” column so that any future doctrine-of-equivalents theory can be blocked at the outset.

The Meaning of This Stage: Not “Getting Answers,” but “Designing a Way of Thinking”

Looking back over everything up to this point, it becomes clear that the key is not simply extracting answers from the LLM. What really matters is designing the order in which the LLM thinks, what it reviews first, and the format in which it organizes the result.

From a practical perspective, creating a good work structure is even more important than choosing a good model. And that work structure is built through the following flow.

1. First, gather the materials.
2. Then identify the gaps in those materials.
3. Supplement the missing pieces.
4. Compress only the core principles.
5. Structure those principles into a framework and workflow.
6. Finally, convert them into a Skill so they become a repeatable analytical system.

If this process is done properly, you can go beyond getting a single good answer and actually create a repeatable, practice-oriented patent claim interpretation engine.


Closing Thoughts

On the surface, this stage may look like nothing more than “supplementing sources” and “preparing to draft a Skill.” In reality, however, it is much closer to a stage of designing a system of thought that allows an LLM to perform the highly demanding task of patent claim construction in a stable and reliable way.

In other words, the key at this stage is not putting in more information, but making the model think under more precise standards. Ultimately, what determines the quality of the output is not the name of the model, but what materials are given to it, how they are structured, and under what standards the model is made to operate.

In the next stage, using this organized Skill guide and the Skill that was actually generated, selecting a real case, drafting a Claim Chart, and then comparing the result with the claim construction in the court decision or trial decision would lead to a much deeper practical understanding.

Mastering Patent Claim Construction with LLMs (3): Injecting Guidelines for Claim Chart Drafting

LLM으로 완벽한 특허청구범위 해석하기 (3)
Practical Framework Part 3

Mastering Patent Claim Construction with LLMs (3)

Injecting Claim Chart Drafting Guidelines Drawn from Practical Experience

The next step is a research phase aimed at establishing the format of the final deliverable.

Rather than organizing claim construction results into a simple narrative report, I prefer to structure them in the form of a Claim Chart specifically designed for claim interpretation. This is not merely to make the output look polished, but to secure a practical tool that can be used under a consistent standard in all downstream tasks, including infringement analysis, invalidity review, and opinion drafting.

1. Why Claim Chart Standards Must Be Established First

There is one important point here. A Claim Chart is not just a “format for organizing results,” but a core framework that determines the quality of the interpretation itself.

The problem, however, is that the method for decomposing claims, the standards for extracting elements, and the actual drafting approach are not systematically organized in textbooks or case law. In other words, this is an area where methodology is built largely 👉 through practical experience.

Through handling a wide range of cases, I have gradually developed my own standards for claim decomposition, element extraction, and chart structure in a way that is suitable for claim construction.

2. Teaching the LLM the “Drafting Method” First

At this stage, what I do is simple. I organize the standards I use in practice 👉 the way I would teach a junior associate, and then register them as a source in NotebookLM.

A Critically Important Point

This step is not simply about “adding reference materials.” It is the process of pre-training the LLM on the interpretive framework and the output format so that it will reason under the same standards in all later interpretation tasks.

  • This is the step where you inject, before adding any source materials, how the analysis should be conducted and how the results should be organized.
  • If this step is skipped, the LLM will analyze each task under a different set of standards, which in turn prevents the overall work product from losing consistency.

3. Claim Chart Drafting Guidelines (Draft)

Below is a draft of the practical Claim Chart drafting guidelines that I prepared and asked the model to review.

DRAFT INSTRUCTION
**1. Methods for Claim Decomposition and Extraction for a Claim Chart Used in Claim Construction** Claims should be broken down into the smallest units that serve as the basis for invalidity and infringement analysis. The language of the claim should be separated according to context, while also being divided into sub-elements in a way that clearly reveals the invention’s distinctive features. The specific methods of decomposition and extraction are as follows. * **Structure- and Function-Based Decomposition:** Break the claim down into **structure**, which reflects the physical form of the elements and how they are connected, and **function**, which reflects the operation and role performed by each element. * **Identifying Organic Relationships:** Go beyond merely listing the components and derive the relationships showing how each component connects with the others and contributes to solving the technical problem. * **Extracting Interpretive Terms and Limitations:** From among the separated sub-elements, extract as the key interpretive targets those terms that **specifically require claim construction or are directly tied to the invention’s distinctive features**. In particular, major extraction targets include terms expressed functionally (for example, “means for ~”) that may need to be examined to determine whether limiting interpretation to the embodiments is required, as well as terms to which the principle of claim differentiation may apply because of dependent claims. **2. The Most Desirable Methodology for Drafting a Claim Chart** The most practical and desirable output format for visualizing claim construction results in a logical and clear manner is to prepare a **Claim Chart** in tabular form. For that purpose, I propose the following table format composed of **six items (columns)** as the output format. | No. | Claim Language | Claim Element (Interpretive Target) | Specification / Drawing Support | Interpretation Result and Legal Analysis (from the PHOSITA Perspective) | Notes (Prosecution History / Limiting Construction) | | :-- | :--- | :--- | :--- | :--- | :--- | | 1 | Decomposed claim language | (e.g., screw) | [paragraph number], [figure number] | (e.g., a low-speed pressing member having a spiral structure) | (e.g., vertical-type limitation based on an argument submission) | **[Drafting Guide and Explanation for Each Column]** * **Claim Language:** Set out verbatim the minimum unit of claim language as decomposed in a way that clearly reveals the invention’s distinctive features. * **Claim Element (Interpretive Target):** Extract from the decomposed language the core target terms or limitations that especially require interpretation or are written as functional expressions (e.g., “means for ~”). * **Specification / Drawing Support (Mapping of Intrinsic Evidence):** Identify and map the specific [paragraph numbers] and [figure numbers] in the detailed description of the invention that define or support the meaning of the term. * **Interpretation Result and Legal Analysis:** From the perspective of a person having ordinary skill in the art, apply intrinsic evidence first and clearly state the objectively and substantively established technical meaning of the claim element. * **Notes (Review of Prosecution History and Limiting Construction):** Record legal points of note, such as any intentional exclusion (prosecution history estoppel) reflected in the submission of arguments or amendments during prosecution, or any reason why, due to a clear disavowal of claim scope in the specification, limiting construction to the embodiments should apply.

This type of Claim Chart format allows intuitive one-to-one textual mapping between the relevant product (or prior art) and the claim elements in later infringement litigation or invalidity proceedings, making it the most effective practical analytical tool for preventing errors caused by improper narrowing or overbroad claim interpretation.

4. The Process of Validation → Refinement → Redesign

Rather than simply using the above guidelines as-is, I validated them in the following way.

“Are these guidelines sufficiently valid under actual legal doctrine and case law standards?”

To answer that question, I instructed the model in the chat to reflect the points requiring refinement and present again the most appropriate concrete Claim Chart format and drafting guide.

5. Refined Claim Chart Drafting Guidelines (Final Version)

The result is as follows, and this too was saved as a memo and then converted back into a source and integrated into the knowledge base. This process of turning good results into knowledge that can be continuously reused rather than a one-off answer is extremely important.

FINAL FRAMEWORK
**1. Methods for Claim Decomposition and Extraction for a Claim Chart Used in Claim Construction (Reflecting Special Issues)** When decomposing claims to apply the All Elements Rule (AER), which is central to infringement and invalidity analysis, it is essential to go beyond a simple listing of claim language and instead perform a more multidimensional extraction that reflects the legal characteristics of **special types of claims (functional claims, PbP claims, and numerical limitation claims)**. * **Identifying Structure, Function, and Organic Relationships:** Break the invention down into the physical structures that make it up and the functions those structures perform, while clearly deriving the relationships showing that each element does not exist in isolation but is organically connected with other elements to solve the technical problem. * **Extracting Targets for Limiting Construction of Functional Claims:** Because elements expressed in abstract or functional terms, such as “means for ~” or “~ step,” carry a risk of making the claim scope unduly broad, they should be extracted as targets for focused review to determine whether there are clearly unreasonable circumstances requiring the claim to be limited to the embodiments disclosed in the specification. * **Extracting the “Structure / Properties” of PbP (Product-by-Process) Claims:** Even when a manufacturing process is recited in the claim, the subject of the invention is still a “product,” so the manufacturing process itself should not be mechanically extracted as an independent element. Instead, the true technical element to be extracted is the “structure or properties of the product” ultimately defined by that manufacturing process. * **Extracting the “Critical Significance” of Numerical Limitation / Parameter Inventions:** Where a specific numerical range or a newly created parameter is involved, it should be separated as a core interpretive target in order to assess whether the specification enables the full claimed range without undue experimentation (support requirement), and whether the claimed range or parameter shows a remarkable effect (critical significance) that distinguishes it from the prior art. **2. The Most Desirable Concrete Claim Chart Format (Top-Tier Practical Template)** By upgrading the existing six-column structure, I propose a **seven-column framework that can fully map both Doctrine of Equivalents (DOE) defense / attack logic in cases of literal non-infringement and the analytical guidance for special types of claims**. | No. | Decomposed Claim Language (Including Organic Relationships) | Core Interpretive Target (Special Issue: Functional / PbP / Numerical) | Specification / Drawing Support (Mapping of Intrinsic Evidence) | Objective Technical Meaning and Result of Literal Interpretation (from the PHOSITA Perspective) | DOE Comparison: Identification of the Principle for Solving the Problem | Defense Logic: Prosecution History Estoppel and Grounds for Limiting Construction | | :-- | :--- | :--- | :--- | :--- | :--- | :--- | | 1 | (e.g., a cleaning unit that supplies cleaning water by electrolyzing filtered water) | cleaning unit (functional claim) | [paragraph 0045], [figure 3] | (e.g., an internal module that performs physical cleaning through electrodes without sterilizing chemicals) | (e.g., the principle of maximizing eco-friendly cleaning efficiency by excluding chemicals) | (e.g., conscious exclusion of a “chemical additive method” during argument submission) | | 2 | (e.g., a tablet manufactured by direct compression) | manufactured by direct compression (PbP claim) | [paragraph 0022] | (e.g., a porous tablet structure having a 15% inter-particle porosity formed by the direct compression process) | (e.g., the principle of controlling disintegration speed by adjusting porosity) | (e.g., structurally / physically different from tablets made by wet granulation) | | 3 | (e.g., a lens assembly having TTL ≤ 6.5 mm) | TTL ≤ 6.5 mm (numerical limitation) | [paragraph 0110], [Experimental Example 2] | (e.g., the physical limit of a miniaturized lens with a total track length of 6.5 mm or less) | (e.g., the principle of achieving an ultra-short focal length through refractive-index redistribution) | (e.g., target for lack of enablement across the full range of 6.5 mm or less) | **3. [Drafting Guide and Explanation for Each Column]** * **Decomposed Claim Language:** Set out verbatim the claim language decomposed into the minimum unit defining the scope of the invention, while describing it in a way that reveals not just a simple list but also the physical and functional relationships among the elements. * **Core Interpretive Target (with Special-Issue Label):** From the decomposed language, extract the key terms that present issues—such as functional expressions (“means for ~”), product-by-process (PbP) language, or numerical limitations / parameters—and identify the type in parentheses so the focus of the analysis is clear. * **Specification / Drawing Support (Mapping of Intrinsic Evidence):** Map the specific [paragraph numbers] and [figure numbers] in the specification that show where the term is defined under the lexicographer rule, where embodiments support a functional claim, where a manufacturing process in a PbP claim causes a specific change in physical properties, or where the technical threshold for a numerical limitation is demonstrated. * **Objective Technical Meaning and Result of Literal Interpretation (from the PHOSITA Perspective):** From the perspective of a person having ordinary skill in the art, apply the principle of construing the claim in light of the specification and describe the objectively and reasonably derived substantive meaning of the element. In the case of PbP claims, the description should not merely repeat the manufacturing process itself, but should clearly translate the structure or properties defined by that process. * **DOE Comparison: Identification of the Principle for Solving the Problem (New):** This column is used to assess possible infringement under the Doctrine of Equivalents when differences are found in a literal one-to-one comparison with the accused product. By considering both the content of the specification and the prior art existing at the time of filing, this column states in advance the “core of the technical idea underlying the specific means for solving the problem unique to the patented invention” (i.e., the principle for solving the problem) to which the relevant element belongs. * **Defense Logic: Prosecution History Estoppel and Grounds for Limiting Construction (New / Refined):** Record the history of any intentional exclusion (prosecution history estoppel) of specific subject matter reflected in arguments or amendments submitted throughout the prosecution of the original application and any divisional applications. Also record whether a functional claim should be limited to a specific embodiment because a broader reading would be clearly unreasonable in light of the specification, or whether a numerical limitation invention may be invalid for failure to satisfy disclosure requirements (such as enablement) because the full claimed range cannot be practiced without undue experimentation.

6. The Core Meaning of This Stage

The most important insight to gain at this stage is the following.

“A Claim Chart is not merely an organizational tool, but a structural mechanism that controls the accuracy of claim interpretation.”

And when using LLMs, the key is not simply getting good results, but building a system in which the results are always generated under the same standards.

7. Key Points for Practical Application

When this process is applied directly in practice, it can be summarized as follows.

  1. First, establish the “drafting format” (the structure of the table).
  2. Next, inject the “interpretive standards” (legal doctrine and drafting guidelines).
  3. Then, input the “source materials” (the specification and evidence).
  4. Finally, carry out the “claim interpretation and chart drafting”.

Most people reverse this order. Because they start by asking for the interpretation right away, the results become unstable and lose consistency.

This stage is not simply about creating a Claim Chart, but about designing how the LLM will think. When this work is done properly, all later analytical work becomes far more stable.

© 2026 ChinSu Lee. All rights reserved.

청구항 해석에 합의했는데 왜 독이 되었나 — Allergan·AstraZeneca 판례로 읽는 합의의 명암

청구항 해석에 합의하고도 불명확성이나 비자명성 다툼은 남을 수 있다. 해석 문제가 모두 끝난 듯 보여도 안심하기는 이르다. 합의 문구가 침해 여부를 가리는 기준이 되고, 항소법원이 판단을 다시 살피는 범위까지 바꾸기 때문이다. 무엇을 끝냈는지...