Are You Tired of Proofreading a Robot?
AI work product has a natural power and allure. It is fluent and plausible, but by now most are familiar with the dangerous practice of treating the chat interface like some black-box oracle rather than interrogating what is going on underneath. Fact-check or die, as some have said.
Even lawyers can fall into over-reliance. The “ChatGPT moment” of November 2022 is now almost four years behind us, and yet law firm blogs still publish roundups of cases where attorneys are sanctioned for relying on hallucinations.
The problem also extends to representation before the USPTO. The Office has issued formal guidance that existing rules (§ 11.18 among them) govern AI-assisted submissions, and that the duty of reasonable inquiry is not delegable to a machine.
While there is no substitute for verifying AI outputs, you can make your job on that score much easier. I am a patent attorney and an engineer, and I got tired of proofreading output that rarely (if ever) says “I am not sure.” So I built a system that forces agentic AI to show its work and verify its sources before it can call anything done. And then I built the tooling backbone that system runs on, called PocketPart, that gives agentic AI the capability to discover and verify facts from primary sources, not web search.
This is an architecture every lawyer who uses agentic AI should build for themselves or else adopt in their practice.
In this piece, I am going to show you how to develop the “skills” suite that forces any agentic AI (Claude, ChatGPT, etc.) to brainstorm, plan, and think about legal workflows the way that you do, to use primary source material to support both fact and argument, and to fact-check its output before returning a final answer.
A suite of skills
A good skills suite is like an operating system for your practice. And it should absolutely be customized (I will show you how). But before we begin, a few ground rules. A good skill (or collection of skills) must include:
The job: or what the task is, its goal and scope, framed in the same way as you would frame it to your junior associate or colleague.
The expected output: what a correct deliverable looks like, including structure, style, and tone.
The failure cases: specific ways the work can go wrong.
The evidence used to judge quality: the primary sources the task must rest on, as well as the standard the work product is held to.
This framework addresses the structural gap that leads agentic AI down the wrong path, causing it to hallucinate case law, rely on stale information, or apply an incorrect idea of what it means to “verify” a fact.
I have found the writing ability of AI is not the main issue (in fact, it is often so good I think lawyers mistakenly over-rely on it). The main bottleneck in our profession is trust, and it is what your skills suite should address.
But what is a “skill” and how do you use it?
We all know how to prompt an AI chatbot. You just type a query and hit return. A skill is like a reusable prompt. It comprises a set of self-contained instructions, kept in a file that the assistant loads whenever it recognizes a that prompt matches the skill. The skill then tells the model how a particular kind of query should be answered. A good skill includes the steps to take, the standard to meet, the sources to consult, and the mistakes to avoid.
Most people do not manually prompt AI in this detailed manner. Let’s be honest for. second, we are too busy (or too lazy) to think deeply about prompting for every chat session. If this sounds like you, adopt skills now.
Skills let you offload mental work (read: be lazy), by saving a workflow that is consistent and correct every single time. Write a skill and you have given the agent a procedure to run every time, the same way, without you needing to remembering it even exists.
Skills, when used correctly, are powerful.
A “suite” of skills is a set, organized to cover the work you actually do, and a good suite behaves like a stack of prompts that automatically get appended over time as the agent works through your problem. As your skills are invoked, they steer the agent through a process you would use yourself if you were answering the question.
So let us dive into the architecture.
A working skill suite has four parts, which mirror how most people work through a problem.
A router: you approach different problems using different strategies, and so should AI. The router is a roadmap that reads the task and loads the right skill before the model does anything else.
When I need to respond to a rejection from the patent office, I use a certain subset of my training. When I am redlining a licensing agreement, I use another subset. The router does this job for the AI agent. Without a router, AI treats prompts as an open-ended writing task, causing it to behave unpredictably.
The process skills: these set the posture and approach. How to brainstorm (a workflow not inherent to AI systems). How to turn a strategy into a plan. And more specifically, how to diagnose certain problems, like why patent prosecution has stalled.
The domain skills: these execute a specific work product. Draft a claim set. Respond to the rejection. Clear the USPTO disclosure obligation.
Think about it. You likely have different approaches for brainstorming, strategizing, planning, and executing, sometimes based on the task at hand. The process/domain skill dichotomy mirrors this explicitly for the agent. A process skill decides how to think about a problem before answering it. A domain skill takes those thoughts and executes to produce a coherent result. You will build many more domain skills than process skills, but the process skills are what keep the domain skills from starting from the wrong place.
A shared verification system: global standards applied before the agent returns the answer: (1) an accuracy check (verification, with a required appendix), (2) a quality check (a rubric, applied by a siloed three-reader review), (3) a definition of done (a checklist the agent must complete or explain), and (4) a volatile-authorities list (the dated facts, kept in one place and marked verify-live).
This is the critical category that makes the result trustworthy. The prior three skill groups will undoubtedly improve your work, but improvement is not the same as trust. If you want to catch hallucinations and stale authority before they reach you, you need a way to verify the output of your domain skills before the agent can call them done:
Most attorneys will inherently understand that accuracy (“is every fact true?” or “is every citation correct?”) is a different question from quality (“is the work good?”). In my practice, I usually review final work product twice: once to fact-check and once using my style guide for quality. I find that I cannot do both at the same time. They require separate “headspaces”; and likewise, I have anecdotally observed that, in agentic systems, the accuracy and quality tests perform best when they are separate and siloed from one another.
In my process, the accuracy gate is an operationalized form of the USPTO’s § 11.18 reasonable-inquiry duty. The skill refuses to let the model skip this step, and if it cannot confirm a fact, it must say so rather than paper it over. My version also requires a verification appendix on all work product, which is a table listing every factual assertion and the source it was checked against. I can read the inquiry instead of trusting that it happened.
The quality gate has two parts: a rubric that defines what good work looks like, and a method for applying it. The method is a three-reader review. For me, an attorney who works more on the patent procurement side, the readers are: (1) the patent examiner, (2) opposing counsel trying to get around or invalidate the patent in ten years, and (3) my client’s general counsel. I have the skill dispatch subagents siloed from one another and from the agent that drafted the work, each reporting back to the orchestrator. Siloing each reviewer prevents cross-contamination. A critic handed only the finished paper is usually more truthful than one exposed to all of the decisions leading into the draft. You should want that criticism.
The definition of done is the checklist the agent must complete before anything is final. If it cannot complete an item, it does not quietly drop it. The agent highlights it and explains why.
The last piece is what I call a “volatile authorities” list. It is a file that holds dated or moving facts, for example: USPTO fees change annually, sometimes even more often; pilot programs sunset, new ones come online; case law is narrowed or distinguished. Each item in the list is marked “verify live before relying.” Keeping the fragile facts in one place, rather than scattered across a dozen skills, helps keep individual parts of the suite from going stale.
Before turning to building, please note: a set of skills does not make the AI model a lawyer, and that is not our purpose here. What we seek is to make the model’s output safer for a lawyer to sign, although you should always personally check the results. Remember, § 11.18 has no AI exception. Do not become another bullet point in a law firm’s AI hallucination blog post.
Building it
You should own your (or your firm’s) legal AI skills.
If your firm mandates a suite, learn how it works. A prepackaged skill hands you another lawyer’s judgement in a form that is difficult to inspect from the output alone. You can read the work product and still not understand what standard produced it, which facts were checked, or what the skill was told to ignore. Either read the file describing how the skill works or ask your AI / IT lead to walk you through it.
Otherwise, if your firm allows it, your time is better spent building your own suite from the framework I described. You will gain a solid understanding of how the suite works and ensure it conforms to the framework and processes you specify.
To help get you started, I will provide a skill suite scaffold below (Github link and single-file bundle). It includes templates for the router, process and domain skills, and the verification gate, along with relevant context for an AI agent.
I suggest you upload the single file bundle and this post into Claude or ChatGPT. Describe your role (attorney, engineer, designer, etc.), the types of tasks you commonly encounter, and ask the agent to help you prepare a skill suite to help you produce stylistically good and factually accurate work product using your own analytical process. I would suggest that your prompt also include a request to “interview” you; this mini-hack will help the model get much better context.
Single file bundle: https://github.com/stnagy/skills-scaffold/blob/main/SCAFFOLD-SINGLE-FILE.md
GitHub: https://github.com/stnagy/skills-scaffold/releases/tag/v0.1.0-alpha.
If you would like to start smaller, begin with just the verification gate; in my humble opinion, this will be the largest initial return on your time investment. I will also paste a sanitized version of my verification skill below, included in the GitHub repo. Note, the skill relies on PocketPart. In the spirit of full disclosure, that is tooling I own; however, several weeks ago, I wrote a post on how you could build your own version: https://stnagy.substack.com/p/im-a-lawyer-i-built-my-own-ai-tools.
Direct link to the verification skill: https://github.com/stnagy/skills-scaffold/blob/main/spine/verification-before-filing.md.
Maintenance and Ethics
Before I close, a note on two items to keep track of.
First, skills become outdated quietly (e.g., as the law changes or procedure shifts), and a static skill will not throw an error when it becomes stale. The AI will continue to produce confident, well-formatted, out-of-date output (which, it should go without saying, is one danger the skill is meant to avoid).
So please maintain your skills. Some thoughts on that end:
Put most hard-coded authorities/facts in one place (e.g., the “volatile authorities” list I described above), so that a change is one edit and not a hunt across a dozen skills. Once a quarter, read that file against current sources and fix what moved.
Require the agent to re-verify entries on the volatile list on each execution. Do not instruct your skills to treat them as the authoritative source of truth. Your skills should use these as a pointer to the current authority, which it should verify live.
Fees change a lot. I do not hard code these. If I think the agent needs a pointer, I cite the relevant CFR section containing the current fee or the location of the live fee schedule.
Second, keep in mind your ethical responsibilities to your client. Nothing about using AI changes the duties you already owe, although it may make it easier in some cases to violate them.
Do not put client-confidential information into a consumer AI tool that trains on your inputs. Use an enterprise or no-train configuration, or do not put it in at all.
Your duty of competence includes understanding the tool well enough to catch its failures. Check the AI hallucination blog I linked above; “the AI wrote it” will not work as a defense.
Every paper you file at the USPTO carries your § 11.18 certification, formed after “an inquiry reasonable under the circumstances.” The model does not make that inquiry. You do. The verification gate helps you keep that promise, but the buck stops with you. Trust, but verify.
AI work product indeed has a natural power and allure. The models are becoming more and more articulate, and they can draft faster than a human ever could. But eloquence without diligence won’t work in the legal industry.
Whatever agent you use, make sure it runs your workflows when it brainstorms, plans, researches, and drafts. It should approach legal work the way you do by grounding every fact and every argument in primary source material, and checking its own output before it tells you the work is done.
Author Note
Stephen G. Nagy is a patent attorney and engineer at Strain PLLC. This article is for general informational and educational purposes only, is not legal advice, and does not create an attorney-client relationship. It reflects the author’s views, not necessarily those of Strain PLLC. Legal authorities change; verify currency before relying on anything here. For advice on your specific situation, consult qualified counsel.
1 https://www.sternekessler.com/news-insights/insights/ai-ip-year-in-reviewai-hallucinations-in-court-filings-and-orders-a-2025-review-of-sanctions-across-the-courts-and-rule-proposals/
2 https://foiadocuments.uspto.gov/oed/Matos-Final-Order-(D2025-13)-Redacted.pdf. Although the facts do not expressly find AI was used, briefing that relies on nonexistent case law suggests AI use.
3 Guidance on Use of Artificial Intelligence-Based Tools in Practice Before the United States Patent and Trademark Office, 89 Fed. Reg. 25,609 (Apr. 11, 2024).
