AI That Reads the Technical Meaning of Patents, Not Just Their Words — Training on Claims and Descriptions to Detect Patent Risk

INVENTOR'S TECHNOLOGY COLUMN · AI AND PATENT ANALYTICS

Training AI on Patent Claims and Detailed Descriptions to Identify Patent Risk

The same technology can be described in entirely different words. This article explains an AI-based patent-analysis technique that turns the native structure of patent documents into training signals and identifies passages in a lengthy disclosure that are most closely related to a claim.

Inventor's note. This article presents, for a general audience, the core idea and operating principles of an AI-based patent-analysis technology that I invented and filed for patent protection. The model's semantic-similarity score is not a legal determination of infringement. It is best understood as an analytical signal for identifying patents and product materials that warrant closer review.

Conceptual overview of AI learning the semantic correspondence between patent claims and the detailed description

Contents

1. Why keyword search misses technical relationships
2. Turning patent documents into training data
3. Working around the 512-token limit
4. How BERT reads the relationship between two texts
5. What the reported experiment shows
6. The line between semantic similarity and infringement
7. Practical applications and the value of data
8. Remaining challenges and the path forward

1. Why Keyword Search Misses Technical Relationships

Patent searching is not simply a matter of finding identical words. Comparing a product description with a claim is relatively straightforward when both use the same terminology. The harder cases arise when two documents describe the same technology in different language. A drafter may reorder components, move between broader and narrower concepts, or describe the same function in different terms. A keyword search can easily miss those connections.

The technology I invented and filed for patent protection begins with that problem. Its objective is to move beyond word matching and train an AI model to evaluate how closely the technical meaning of a claim corresponds to a passage in another document. The central idea is to use structural relationships already present within patent documents as training data. Rather than requiring experts to read tens of thousands of documents and manually create an answer key, the system is designed so that the patent documents themselves supply part of the training signal.

2. Turning Patent Documents into Training Data

The starting point is the structure of a patent document. The claims define the legal boundaries of the invention, while the detailed description explains what the invention is and how it may be implemented. Article 42(4)(i) of the Korean Patent Act requires the claims to be supported by the detailed description. That requirement helps prevent an applicant from monopolizing subject matter that was not disclosed and enables a person skilled in the art to understand the relationship between the claimed invention and the specification.

Viewed through a machine-learning lens, that legal structure creates an opportunity. A claim and a passage from the detailed description of the same patent will ordinarily have a strong relationship. By contrast, a claim and a randomly selected passage from a different patent are statistically more likely to be unrelated. The system can therefore assign label 1 to a claim–description pair drawn from the same patent and label 0 to a pair assembled from different patents.

An essential distinction. Label 1 does not mean “legal infringement,” and label 0 does not mean “legal noninfringement.” These are proxy labels designed to teach the model semantic correspondence, not ground-truth labels derived from judicial infringement decisions. Two different patents may concern related technology, and a claim and description from the same patent say nothing by themselves about whether a third party infringes.

The immediate benefit is a substantial reduction in manual labeling. In a conventional supervised-learning project, experts might need to read each document pair and decide whether it is related. Patents take time to read carefully, so the cost rises quickly as the dataset grows. Automatically pairing claims with passages from the same patent—and combining claims with randomly selected passages from other patents—can remove much of that bottleneck.

Negative samples are not mere filler. If a model sees only highly related pairs, it may learn to treat a few overlapping words as sufficient evidence of a meaningful relationship. Training it on claim–description pairs drawn from different patents forces the model to distinguish superficial word overlap from technical correspondence. Positive samples show the model what a relationship looks like; negative samples help define its boundary.

3. Working Around the 512-Token Limit

Document length presents the next obstacle. In its basic form, BERT accepts no more than 512 tokens in a single input. A patent's detailed description, however, often runs to thousands of words. Feeding a claim and the entire description into the model at once is therefore impractical.

My proposed approach first divides the detailed description into chunks of approximately 310 tokens. It then calculates the semantic relationship between the claim and each chunk and selects the highest-scoring passages as input candidates. This differs from mechanically taking text from the beginning of the document. The limited input window is allocated first to passages most likely to bear directly on the claim.

A dot product between vector representations can be used to rank the chunks. The underlying intuition is straightforward. The claim is represented as one semantic vector and each description chunk as another. The operation scores how closely the two vectors point in the same direction. A higher score indicates a greater likelihood that the claim and the passage are semantically related.

1. Segment: Divide the detailed description into chunks of roughly 310 tokens.
2. Score: Calculate the semantic relationship between the claim and each chunk.
3. Select: Choose the highest-scoring chunks as candidates for the BERT input.

4. How BERT Reads the Relationship Between Two Texts

The selected text is formatted so that BERT can distinguish the two inputs. A representative sequence is [CLS] + claim + [SEP] + detailed description + [SEP]. The [SEP] tokens mark the boundary between the texts, while segment embeddings identify which tokens belong to which input. The final representation of the [CLS] token can then serve as a summary vector for classifying the relationship between the two texts.

This structure prevents the claim and the description from collapsing into a single undifferentiated passage. Human readers naturally use headings, paragraphs, and line breaks to identify document boundaries. A model has no comparable intuition unless the input structure expressly tells it, in effect, “Text A ends here, and Text B begins here.”

During training, label-1 and label-0 examples are handled through an alternating-batch strategy. Rather than mixing both classes indiscriminately from the outset, the system calculates the loss for each group and updates the model parameters in a direction that reduces both losses. The purpose is to keep the model's decision rule from drifting toward one class merely because that class is larger or easier to learn.

5. What the Reported Experiment Shows

In the experiment described in the source manuscript, cross-entropy loss decreased from 0.89 to 0.19 over 15 training epochs, and accuracy reached approximately 80%. The manuscript also reports that the model classified all 17 text pairs in a comparison set that used different terminology as label 1.

15 epochsTraining duration

0.89 → 0.19Cross-entropy loss

Approx. 80%Accuracy under the reported conditions

How to read the numbers. These results were obtained under the specific experimental conditions reported in the source manuscript. They do not establish that the same performance will hold across every industry, patent corpus, or dispute. The result for the 17 text pairs should likewise be understood as an observation from a limited comparison set.

More important than the headline numbers is the model's response to changes in wording. One of the central difficulties in patent searching is that the same technology can be obscured by different language. Suppose one document refers to an “adhesive retention mechanism,” while another describes the same arrangement using entirely different terminology. A keyword search may push the documents apart because they share few words. A semantic model is intended to do the opposite: it asks what function each component performs in context and how the components relate to one another.

6. The Line Between Semantic Similarity and Infringement

This discussion naturally calls to mind the doctrine of equivalents. Patent infringement analysis does not necessarily end merely because an accused product departs from the literal wording of a claim. Under Korean Supreme Court precedent, an altered element may in appropriate circumstances be treated as equivalent when requirements concerning the invention's problem-solving principle, substantially identical operation and effect, and interchangeability are satisfied, absent a recognized bar to equivalence.

An AI score is not an infringement ruling. Semantic similarity and infringement under the doctrine of equivalents are not the same concept. A high similarity score does not establish infringement, and a low score does not eliminate legal risk.

A legal analysis must separately address whether every claim limitation is satisfied, as well as prosecution history, the prior art, any deliberate exclusion or disclaimer, and each requirement of the doctrine of equivalents. The practical value of this invention therefore does not lie in having AI replace a judge or lawyer. Its more useful role is to narrow the set of documents that a human expert must read. Among thousands—or tens of thousands—of patents and product documents, the system can elevate candidates that correspond strongly to a claim, after which patent counsel, litigators, and R&D personnel can conduct a rigorous review.

7. Practical Applications and the Value of Data

The same semantic-comparison architecture supports several practical use cases. Patent-risk analysis, portfolio monitoring, and prior-art searching all converge on the task of finding technical relationships expressed in different language.

• Product development: Compare competitor patents with product specifications to identify potential risk candidates early.
• Rights monitoring: Compare descriptions of newly released products with existing claims to prioritize human review.
• Prior-art searching: Surface references that describe relevant subject matter using language different from the claim.

In patent AI, the most expensive resource is not always GPU capacity or model parameters. Often the harder problem is obtaining good training data. Expert labels can be accurate, but they are slow and costly to produce. Public patent documents are abundant, yet they are not ready-made answer keys. This invention seeks a middle path: use the legal and documentary structure already embedded in patents to reduce the cost of generating training data.

Describing the project simply as “turning law into mathematics” misses the point. More precisely, it identifies signals that AI can learn from the structure and rules accumulated in legal documents. The relationship between claims and the detailed description, the selection of relevant passages, the construction of positive and negative samples, and the formatting that distinguishes two text inputs together convert semantic correspondence in patent documents into a computationally tractable problem.

8. Remaining Challenges and the Path Forward

Substantial challenges remain. Terminology and sentence structure vary by technical field, which may require domain-specific training. Random negative sampling can also produce false negatives by pairing documents that are, in fact, technically related. Narrowing the gap between semantic similarity and legal infringement analysis will require claim-element-level analysis supplemented by prosecution history, prior art, and expert-labeled data. BERT's 512-token constraint may also be addressed through longer-context models or hierarchical architectures.

Even so, the starting proposition is clear: a patent is not merely an undifferentiated block of text. A special relationship exists between the claims that define the legal right and the detailed description that explains the technology. Once that relationship is converted into training data, AI can move beyond retrieving documents that use the same words and begin exploring why two documents may be technically related.

The proper role of AI in patent practice is not to displace the expert's ultimate judgment. It is to find, more quickly and across a wider field, the candidates that deserve that judgment. The system first detects semantic connections that may be easy to miss in a large corpus; lawyers and technical specialists then apply the governing legal and engineering standards. The invention I filed occupies that boundary. It is an effort to move beyond searching the words of a patent and toward reading the technical meaning the patent seeks to protect.

Reference

ChinSu Lee (이진수), “Method for Generating a Patent-Analysis Model Using an Artificial Neural Network and Text-Pair Embeddings, Patent-Analysis Method, and Computing Device” [English translation of title], Korean Patent Application No. 10-2024-0075102, filed June 10, 2024.

• Application number: 10-2024-0075102
• Filing date: June 10, 2024
• Priority application: 10-2023-0093439 (July 18, 2023)
• Inventor / applicant: ChinSu Lee (이진수)
• Original Korean title: 인공신경망 및 텍스트 쌍 임베딩을 이용한 특허 분석 모델의 생성 방법, 특허 분석 방법 및 컴퓨팅 장치

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

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

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.