ChatGPT Isn’t Broken — Your Workflow Is. #chatGPT #ai

I've been thinking a lot about how lawyers use ChatGPT, both in my own law firm, with some of the lawyers that I work with day to day, and I keep coming back to one principle. If you don't control the token chain, the token chain ends up controlling you. Most lawyers, unfortunately, they treat ChatGPT like a single input and output system. They basically ask a question and then they hope that they're going to get an answer. That's not how ChatGPT works. If you blindly dump materials in when you ask a question and hope that the model tracks everything, it just simply won't. And when it tries to try and figure out where you're going and what context that you've provided that's relevant, irrelevant, that's when you get into hallucinations, vague answers, or even inconsistent answers. And the loss of precision is going to be The loss of quality in your work product. And this is what we hear lawyers complaining about all the time, right? ChatGPT is not working for me. It's too general. It doesn't give me the results that I'm looking for. It hallucinates. You hear it all the time. No, the problem is not ChatGPT. It's you. You're not strategic in your use. You're not skilled in your use, and you don't have the right mentality when you sit down at your computer, use it. So the solution isn't necessarily better prompts, although certainly better prompts helped. It's really creating a controlled environment. It's what are you pouring into the mix? When you're making soup, you control what ingredients go in, you control when those ingredients go in, you know what the end result is supposed to be, and you want to think of your ChatGPT sessions in a similar way. You have to manage the instructions. You have to manage the knowledge. You have to manage the context. You have to manage the workflow. That's how you get reliable, repeatable, legal-grade output that's going to take you from 2x to 10x to 20x. So today, and at the end of this video, I'm going to give you the real magic behind this. But today we're going to be talking about the workflow that I've been using all day today. Everyday AI in order to get really A+ results. And the, the tips I'm going to share with you today are very simple, but the payoff is going to be huge. I guarantee it. So just break this down into steps. We've talked about a lot of these things before. I'm going to try and put them together into a comprehensive step-by-step process that you can follow. And you should go back and look at some of my prior videos about, for instance, how to set up a project folder. You know, how to go about determining what is a project folder. Today we're going to focus in on the step-by-step approach to getting great results, starting with setting up the project folder. So every matter or every issue gets its own project folder. And in a video I did just a couple of days ago, we talked about how really the, the more narrow the issue, the more important it is to get its own project folder set up. So instead of having a client matter project folder that does everything, You want to have that so you've got this kind of general purpose tool for the client and the matter, but you also need for each client matter task category folders or even task level folders that you're going to control the context, you're going to control the tokens that go into that specific project folder because less is more. It's not only about what you put in, it's what you keep out of the token chain. Token chain is everything that's going into that project folder by way of instructions, by way of uploaded knowledge by way of your input and by way of the output that the GPT gives you. So every matter that you are going to work in a project folder for, you want to put into that folder the core instructions using metaprompting. In my other videos, we've talked about how to do metaprompting. You're going to tell it what you want the model to do. You're going to store the case background. You're going to upload knowledge into that project folder that is going to be not every piece of information that you've received from the client, but just the core pieces of information, edited Word documents that are going to contain just the context that you need for the specific task that you're trying to achieve. So you want to store the case background, the constraints, the roles, and the tone, put that into the instructions. And if you watch some of my other videos, you know how to do markdown files, which will even constrain that project folder even more to only do specific things in specific sequences and to give you a specific format. So you really need to control on a project level what your goal is, what you're going to put in by way of the token chain in order to get the best results. And, but that's not it. That's not the real magic. Stay tuned. We're going to talk about how to take that to the next level. So this foundation that you create, this becomes your anchor for everything downstream. And it's a stable environment the model can rely on. The project folder that you've created, it's got just the right amount of information and just the right meta-prompting instructions in it. Going to create this stable environment for your work product that's going to make it much more reliable, far fewer hallucinations, and get you much better quality results than the ones that you're getting already. So step One, create that project folder and be very mindful about how you do it. Be very strategic about the information that you put into it. Step 2 is to build a clean foundation thread within the project folder. So you've got your project folder. Now you're going to start a thread. And if you have a client intake transcript, an email chain, or a summary that you've already created, you might upload that into the thread. And then you might ask that thread to extract the key facts, identify the key issues, the, you know, the key background information, the risk tolerance of the client, your first impressions, anything that's really high level that all future work in that thread is going to depend on. So I'm not putting the whole case into one thread. I'm building the base foundation within each thread, just like I did with the project. Now I'm getting even more narrow and I'm Creating that token chain. It's going to be the project folder. Now it's going to be the thread, and I'm going to work from that foundation on future work. And, and it's not going in the direction that you think, because what we're going to do— something is very special here with this foundation that we've created within the thread. So this foundation, you want to think about it as the ground truth that you can keep updating and working with As you maneuver through the project. Step 3, you got your project folder, you've started your thread, and you've thought strategically about what goes into the project folder. You've thought strategically about what goes into the thread. So now you're going to do something that you're probably not thinking about, which is to use the branching feature within the thread, and you're going to do it strategically on the token chain foundation that you've built. And this is the part that lawyers Most lawyers aren't using it at all, but it is extremely powerful and I use it all day long, every day. So branching your threads is the power move. So once the foundation thread is clean and stable, I create new branches for specific tasks. Every branch inherits the curated context of the foundation, but it stays focused on a single task. And that separation matters more than people realize because Within each thread, you're going to be able to create a new, much more narrow token chain that focuses just on a specific issue. So here are some types of branches that I use. So a branch for factual development. If I want to follow— if I want to draft follow-up questions for the client based on that initial token chain that I created and then thread from there, I might create a branch dedicated solely to saying, you know, what further information do I need from the client? So the model isn't mixing factual development with legal strategy or research, which are going to be separate threads, and it is able to procure really precise, relevant questions. Now I might take that initial thread, that foundational thread that I created, and I might branch that exact same thread at that exact same point and now start to do legal research. So I got the basic factual information. I might say, let's start to think about what legal claims in a particular state or under a law might be applicable. Let's do some brainstorming. So I did factual development, questions to the client. Now I'm going to do brainstorming to see if we can kind of think about what kind of legal research we should be doing in the future. So if I need case law or statutory analysis, I'll open up that new branch. And the model isn't distracted by my earlier other branches. And it's not one continuing thread where I'm trying to do factual development, questions to the client, emails to the client. on that same thread, I'm going to start to do brainstorming on legal theories. No, separate them out, branch everything into its natural thread. And that way, you're really doing, again, context engineering. You're controlling what goes into the token chain. Less noise means fewer hallucinations and tighter analysis. So let's talk about Another branch that we might do, right? So you've got the foundational threaten. You've branched that off into doing some brainstorming on legal theories, angles, and approaches. You might then go to Westlaw, get some cases to really start to tighten that up and make sure that your brainstorming is actually backed by case law. Keeping in mind that ChatGPT is not a legal research tool. It's a brainstorming tool for legal theories. And you're going to then say, okay, I want to do some drafting. So you're going to go back to that foundational thread, take a look at it and go, okay, that's good. I'm going to go to my second thread on legal research and I might branch off from that thread into a drafting project. In fact, I might take that second thread that's doing legal research or that's brainstorming legal research, And I might branch off 4 or 5 different sections of a brief from that thread. So the first branch of the legal research thread might be, let's do an outline. And then from that outline, you might do another branch that's like, let's do the introduction section. Then we might do another branch from that outline that is going to be, let's go ahead and draft the initial arguments on this particular issue. And you might go and you might say, we want to use Canvas again to do a branch from the outline. on yet another section of the brief. So breaking it up into pieces, controlling the token chain, it's going to be really important. And with each new section of the brief, you'll take that final version of the prior section, put it in Word, and then upload your final work product so it has that context to do the next section. And you probably are going to say things like, avoid duplication, don't re-argue things that have already been covered in the prior section, only expand on things that are in the prior section. You might tell it to use the prior work product as context so that we end up with a cohesive brief at the end of the drafting project. You're going to want to tell it what you're doing. I'm drafting a brief. I'm using threaded discussions. I'm using each thread to draft each section of the brief so that it knows what your approach is, what your strategy is as it does its work. And you might do another branch for risk assessment, and you're going to pick which thread and at what point within that thread you want to branch off to do risk assessment. You might do another, yet another branch for the client communication when you get the final brief to send that draft to the client for review. That's, that's an amazing approach. So why does branching improve output? It's really 2 main reasons. Reason 1, it reduces the cognitive load for the model. Large context windows encourage lawyers to stuff everything into one place, but more context is not always better. In fact, it's typically worse. The model performs best when the context is clean, relevant, and bounded. Branches help protect that. Reason 2, it reduces hallucinations by limiting the pull of unrelated parts of the conversation. One of the reasons why you see hallucination is because you've provided the chat with way too much disparate, irrelevant information, which is still now being analyzed every time you do input and output. It's still got all this irrelevant context that it might pull into the conversation because it doesn't know that it's irrelevant. And you included it in the context by not excluding it from the context. So hallucinations, they spike when the model is trying to blend tasks. If you mix facts, strategy, research, drafting, and speculation into one thread, the model is going to be eventually, it's going to break down. And you'll see that as these threads get longer. So branching helps prevent that. Each task is shielded from the others. So the third reason why this is super helpful is that it helps you avoid something called creeping substitution. Creeping substitution is when the model replaces your earlier context with its own guesses because the thread has drifted. By branching early, you freeze the context before drift happens. So you've got hallucinations and you've got drift. Drift is where— hallucinations is where it just gets it wrong. Drift is where it starts to get off into the weeds, and you may or may not realize that that is happening. So step 4 of your branching strategy is to use branching to scale your mind. Branching just isn't about creating more chats. It's about creating a parallel work stream for each task that feels specialized, and it feels in many ways like a specialized associate working At that task and having multiple associates doing different things at the same time. So you got lots of different paralegals and associates. Do you have one of them do everything, or do you pick each associate and paralegal based on their specialization? Branching is like creating specialized associates and paralegals, empowering them with the right context and specialization and instructions in context. And then turning them loose. So you get one lane for research, one associate for research, one lane for drafting, one lane for factual development, one lane for strategy exploration, another lane for communication. Each lane hopefully stays relatively clean and sharp. This is really what we want to try and keep thinking about as we think about the Quad Core learning method for learning how to best use AI. Remember, you've got mindset. What is the attitude that you bring to the desk when you sit down at your Wrong attitude, you're going to give up, you're going to stray, you're not going to care enough, you're going to use the AI to do shortcuts rather than to do complex thinking. Wrong mentality, you are going to lose in the age of AI. But you also need strategy. Strategy is understanding what your end zone is and then figuring out the steps to get to your end zone and then implementing the right strategy in using the AI tool to break things up into pieces. In components so that you can end up in the end zone. So when you want to upgrade the common foundation and you say you learn something new from the client, you go back to the foundation thread, you update the facts, and you branch again from that new foundation that you've created. These branches are disposable, and the foundations that you create that you keep branching from, you want to think of them as more permanent. As more reliable places to go back and pick up your work. So this branching workflow has changed the way I handle matters at really fundamental levels. It has reduced the noise, eliminated the drift, and eliminated a lot of the hallucinations that you can sometimes see. Most importantly, it kind of mirrors how real-life legal teams actually work in the real world. Each associate or paralegal has a different lane, different tasks, different specialization, different talents, and you create those using these GPTs. threads as you move forward. When lawyers stop treating ChatGPT like a single conversation and start treating it like a structured workspace and use it strategically, everything changes. You get clarity, you get speed, you get reliability, you get better legal work for your clients than you could have ever done before these AI tools were available. My name is Enrico. Go ahead and subscribe. Go ahead and share. Go ahead and ask questions in the comments. We're all in this together. We'll see you next time.

To a world where machines dream, AI whispers secrets, man first sings. Till tomorrow glows bright on the screen, future unfolds, come join the team.