How to Design a Workflow That Turns a Clueless AI into a Patent Expert
As promised in my previous post, this article will provide a concrete guide on how to perform patent claim construction using LLMs.
To be precise, the method I am introducing involves taking the knowledge gathered through Google NotebookLM’s Deep Research feature and injecting it into Claude in the form of a Skill. While there are various approaches out there, my experience shows that this method consistently delivers the most stable and reliable results.
*Note: There is also an MCP (Model Context Protocol) that links NotebookLM with Claude (Code or App). However, since it is currently only distributed unofficially via GitHub, I am not yet using it in my actual practice.
Expansion into Various Patent Analyses and Core Elements
While this article focuses on Claim Construction, this exact workflow can be seamlessly extended to the following tasks:
- Patent Infringement Analysis
- Patentability Search
- Invalidity Search
However, there is a critical element common to all these analyses: how to acquire and inject into the LLM the Common General Knowledge (CGK) and Prior Art held by a Person Having Ordinary Skill in the Art (PHOSITA) in that specific technical field.
This is the core factor that dictates the accuracy of the results, and the specific methodology often comes down to individual know-how. In practice, the following three aspects are crucial, and they must ultimately be designed based on a deep understanding of patent legal principles and case law:
- The temporal scope of the prior art
- Setting the boundaries of the common general knowledge
- Precise selection of the target for investigation
This process is remarkably similar to assigning tasks to a new hire. If you don’t provide them with sufficient background and clear standards, you can’t expect consistent results. The output of an LLM is inherently probabilistic. Therefore, even with the same model, the conclusions can vary drastically depending on the instructions provided. (For reference, under default settings, Claude tends to maintain relatively more consistent responses.)
Ultimately, the Key is “Workflow Design”
In practice, it is highly recommended to follow these principles:
- Break down the task into step-by-step stages.
- Provide clear guidelines for each stage.
- Reflect legal principles and case law standards in the guidelines.
Furthermore, because current commercial LLMs have Context Window limitations, you must adhere to the following:
- Divide the subject of analysis into the “smallest possible units.”
- Execute the work in segmented batches as well.
Ignoring this can cause the model to lose context, leading to a sharp decline in the accuracy of the results.
Practical Workflow: Injecting Legal Principles via Deep Research
Let’s look at the actual workflow. First, create a new notebook in Google’s NotebookLM. You can think of this step as hiring a dedicated analyst. Next, use the Deep Research feature to investigate case law and legal principles related to claim construction, and register these as Sources in NotebookLM.
After activating the “Deep Research” feature in the initial chat window, enter the following instructions:
PROMPT 1
Research patent claim construction rulings issued within the last 5 years by the Supreme Court and Patent Court of Korea, and the Supreme Court and Intellectual Property High Court of Japan.
To obtain more sophisticated results, it is highly recommended to refine your prompt as shown below. This is the exact same approach as providing detailed operational guidelines to a junior associate in a real-world setting. I used the following prompt:
PROMPT 2
Analyze Korean and Japanese case law regarding patent claim construction based on the following conditions.
## Scope of Research
- Timeframe: Within the last 5 years
- Korea: Supreme Court, Patent Court rulings
- Japan: Supreme Court, Intellectual Property High Court (including Tokyo IP Division) rulings
- Target: Precedents where “patent claim construction” is explicitly stated as the main issue
## Selection Criteria
- Rulings that contain legal reasoning, not just simple factual determinations
- Prioritize precedents that include the following issues:
1) Literal interpretation vs. referencing the detailed description
2) Application of the Doctrine of Equivalents (DOE)
3) File wrapper estoppel / Conscious exclusion
4) Interpretation of functional or abstract language
5) Limitation to specific embodiments
## Analysis Items per Precedent
- Case Name / Court / Year of Ruling
- Summary of facts (3~5 lines)
- Main Issue (portion related to claim construction)
- Summary of the holding and legal principles
- Relationship with existing precedents (expanded / maintained / changed)
- Practical implications
## Comparative Analysis
- Differences in interpretation standards between Korea and Japan
- Common legal principles (e.g., principle of literal interpretation)
- Differentiated approaches (e.g., how the DOE is applied)
## Output Format
- Mix of tables and descriptive text
- Summarize key legal principles in bullet points
- Highlight important precedents separately
## Additional Requirements
- Include links to the actual ruling text or official sources for each precedent where possible
- If precedents are scarce, supplement with case commentaries, academic papers, theories, or explanatory materials
A short time after executing the prompt, NotebookLM will compile one summary report and about 20 related documents, asking you whether to add them as sources. At this stage, review the “key sources,” filter out any unnecessary or low-reliability materials, and instruct it to add all the remaining sources.
This process is not just about organizing files; it is a critical step in controlling the quality of the baseline data for your analysis. In practice, you might find that some sources fail to load properly. Because these failed sources can degrade the accuracy of the analysis, it is highly recommended to remove them completely.
*Note: Even if you use the exact same prompt, the results may vary depending on the execution time or the user’s environment. This is a natural phenomenon, as LLMs operate probabilistically.
Extracting and Verifying Core Legal Principles
In the next step, you will extract the common legal principles regarding claim construction based on the collected case law. This is akin to instructing a junior associate to “summarize the research findings and extract the core legal rules.” To do this, input the following prompt:
PROMPT 3
Regarding patent claim construction, what are the core principles and legal rules that commonly appear across all these sources?
When you review the results generated at this stage, you might notice issues: the model often quotes the ruling text verbatim, resulting in abstract standards that are difficult to apply directly in practice, or it might reflect certain legal principles incompletely. (For example, I once found that the principle of referencing prosecution history was mentioned only narrowly in relation to the Doctrine of Equivalents.)
Accordingly, I provided my own separately compiled practical rules of thumb and instructed the model to compare and verify them against its own case law analysis. This is identical to a senior attorney supplementing a junior’s findings and asking for a review. I have been consistently collecting case law and academic papers where patent claim construction was the main issue, feeding them into NotebookLM to synthesize the common legal principles.
The Claim Construction Framework that I have established is as follows.
GUIDELINES (FRAMEWORK)
This is a 5-step construction framework that takes “Claim-Centric Interpretation (Literal Interpretation)” as the grand principle, while “supplementarily referencing the detailed description and drawings” to objectively and reasonably determine the technical significance.
5-Step Framework for Patent Claim Construction in Korean Courts
Step 1 (Principle of Literal Interpretation): The scope of rights is primarily defined by understanding the terms used in the claims based on their ordinary and customary meaning in the relevant technical field.
Step 2 (Referencing the Detailed Description and Drawings): The overall context of the specification is referenced from the perspective of a Person Having Ordinary Skill in the Art (PHOSITA). If the applicant explicitly defined a term in the specification (the lexicographer rule), that definition takes precedence. Specifically, the meaning of the term is interpreted through the detailed description to solidify its technical meaning so that the intended functions and effects of the invention are realized. The Supreme Court strictly requires that when referencing the specification, one must go beyond merely looking at the context and objectively and rationally consider the “technical significance (the principle of solving the problem and its operational effects)” intended to be expressed by that language.
Step 3 (Prohibition of Restrictive/Expansive Interpretation): Even when referencing the specification, the claim scope must not be unreasonably restricted based on specific embodiments, nor should it be forcibly expanded beyond the specification. Specifically, limitations or features of the embodiments described in the detailed description that are not recited in the claims cannot be arbitrarily imported into the claims. The most heavily guarded error in Korean patent litigation practice is confusing “interpreting the meaning in light of the specification” with “importing limitations from the specification into the claims for a restrictive interpretation.” The Supreme Court maintains a firm stance on strictly blocking unjust restrictive interpretations based on the description of the invention (e.g., the Cream case, the Display Structure case).
Step 4 (Referencing Prosecution History and Estoppel): If a specific configuration was consciously excluded or limited during the prosecution process, it is prohibited to later reverse this and interpret the claim expansively (File Wrapper Estoppel).
Step 5 (All Elements Rule and Doctrine of Equivalents): When determining infringement, the All Elements Rule (AER) is strictly applied, but the Doctrine of Equivalents (DOE) is applied complementarily to encompass design-arounds.
After inputting this framework, I requested verification as follows:
PROMPT 4
Based on a close analysis of all sources, I have drafted practical claim construction principles and a framework. Evaluate this.
As a result, I received a positive evaluation regarding its alignment with case law, and at the same time, the model suggested improvements for areas that needed supplementation. This process goes far beyond simple information gathering; it forms the following iterative loop:
👉 Extracting Case-Law Based Principles → Injecting User Knowledge → Re-Verification
Through this structure, you can mitigate the LLM’s hallucination limits while deriving highly objective results that are immediately applicable in actual practice.
(In the next installment, we will continue by covering how to write and refine the guidelines for creating a Claude Skill based on this data.)
