During the COVID-19 pandemic and in the years immediately afterward, I worked with a company developing rapid electronic sensors for detecting viruses. Our work later expanded into the much more difficult problem of developing rapid electronic sensors for sepsis. During that period, we prepared several patent applications, and I made extensive use of proprietary AI-assisted patent tools.
AI was becoming part of our scientific work as well. We used it for bioinformatics, sensor research, and modeling. At the same time, I was teaching bioinformatics, biotechnology, and microbiology. I was therefore seeing artificial intelligence from several different directions—as a research tool, as a teaching tool, and as an aid in preparing patent applications.
When I left those positions, I began working as an AI trainer in bioinformatics. That experience gave me a very different perspective on large language models. We would pose difficult scientific and research problems specifically designed to expose weaknesses in the models. In other words, part of our job was to make the AI fail.
Once we identified where and why a model failed, we developed detailed rubrics showing how the problem should be analyzed and what a correct response should contain. Those rubrics could then be used to help train the AI to handle similar problems more effectively in the future.
That work naturally led me to a question: Could the same approach be applied to patent work?
I had a particular interest in finding out. For nearly forty years, I had worked with inventors from many walks of life, helping them obtain patents while also working on patents of my own. One problem remained remarkably consistent throughout those years: professional patent assistance and professional patent-searching tools are expensive, particularly for an individual inventor.
The proprietary AI-assisted tools we had been using professionally were excellent in many respects, but they were also expensive. Two licenses alone cost well over $1,000 per month. That might be manageable for a company or law firm. It is obviously impractical for an independent inventor working at a kitchen table, in a basement workshop, or in a garage laboratory.
So I began experimenting with the increasingly capable general-purpose large language models available to everyone.
Initially, I was appalled by how poorly they performed on many patent-related tasks.
The problem was not simply that they occasionally made mistakes. They frequently approached patent work in fundamentally inappropriate ways. A model might ignore the actual statutes, regulations, USPTO guidance, or established patent practice and instead construct an answer from whatever material happened to be readily available in its training data. That could include online discussions, blogs, Reddit conversations, and advice written by well-meaning people who had little practical understanding of patent prosecution.
Patent searching was an especially obvious example. I repeatedly encountered AI-generated instructions that treated a few keyword searches in Google Patents as though they constituted a professional prior-art search. They do not. A serious patent search normally goes considerably further, including classification searching using systems such as the Cooperative Patent Classification (CPC) and searching the patent databases maintained by organizations such as the USPTO and European Patent Office.
I began approaching the problem much as I had approached AI training in bioinformatics. Instead of simply asking an LLM to "write a patent application" or "search for prior art," I broke the work into individual tasks. I looked for the ways the models failed at each task and then developed instructions, guardrails, required inputs, checks, and structured workflows intended to prevent those failures.
At the same time, the LLMs themselves were improving rapidly. New generations became better at following lengthy instructions, analyzing uploaded documents, maintaining context, and carrying out complex, multi-stage tasks. Work that had been unreliable or impractical with an earlier model became surprisingly effective with a newer one—provided that the model was given the right instructions and the right source materials.
Gradually, those experiments evolved into a collection of specialized prompts and workflows for patent research and patent preparation.
The objective was never to turn an LLM into a patent attorney or patent agent. Nor was it to ask AI to invent something on behalf of the user. The objective was much more practical: Could we give an independent inventor access to some of the structured methods, analytical discipline, and AI-assisted capabilities that previously required expensive professional software?
That question ultimately led to the prompts that form the basis of this book and the accompanying Prompt Library Pro.
The library represents lessons learned from several different parts of my career: decades of working with patents and inventors, years spent in scientific research and teaching, practical experience using expensive proprietary AI-assisted tools, and later experience deliberately testing and training large language models.
The result is not a single "magic prompt." It is a structured collection of prompts designed to make an LLM slow down, ask for the right information, work from the inventor's actual disclosure, follow established patent procedures, identify missing information rather than invent it, and produce materials that an inventor can use to better understand and document an invention.
In many ways, that brings the project back to where it started. The technology has changed enormously over the course of my career, but the inventor's basic problem has not: How does an individual with a good idea get access to the tools and knowledge needed to give that idea a fair chance?
Prompt Library Pro is my attempt to make one part of that process a little more accessible.