Best practices: AI prompts for lawyers

Greetings, everyone, and welcome to the webinar channel of the Association of Certified eDiscovery Specialists. My name is Mike Quatorero. I am the president of ACESED. Today, we are gonna be joined momentarily by our partner, Opus two, for a webinar entitled ask smarter, prompt building, best practices for law firms. We’ll get into the presentation in a moment. In the meantime, you can visit and learn more about Opus two at opus two dot com on the web. Alright. Before we get started, plea no please know we love questions. We’re happy to take questions. You can submit the questions using the chat feature or the q and a button. If you use the q and a button, you may submit your question anonymously. Also, if you’d like a copy of today’s slide deck, it’ll be available in the resources button located on the bottom of your screen. Without further delay, I am happy, super excited to introduce our presenter, to and hand off to Smriti Sinha, who will take the presentation from here. Smriti, please take it away. Thanks, Mike. Welcome, everybody. Thank you so much for joining us, and a big thank you to ASEDs for hosting us today as well. So as Mike just said and as the slide says, we’re gonna spend a bit of time today talking about effective prompting strategies. Now spoiler alert before we start, I don’t think there is a single best way to prompt. But what there are, there are techniques that you can learn and apply that help you get consistently consistently get better and more accurate results. And there are also legal duties that lawyers using AI must comply with. So we’re gonna cover a bit of both today. We have an hour. We’re not gonna be able to cover all things that are prompting. But, hopefully, what you’ll walk away with is some practical techniques that you can start applying immediately, and start seeing better and more effective results from your AI use. So just two minutes on introductions. So as Mike said, my name is Rishi Sinha. I’m general counsel at Opus. I have been in house for, quite a long time now, but working across many different sectors. A big part of my role at Opus two now is AI governance. Myself and my legal team uses AI every single day as do most other teams across the business. And in my role, I sort of I look at AI in two different lens. So one is responsible use. So how are we making sure that we’re using AI responsibly, safely, and for lawyers in accordance with our professional obligations? And the other part is the practical deployment. So how are we making sure that the AI that we’re using is effective, is giving us good results, and is giving us results that we can stand behind? I’m joined by my colleague, James. I’ll let James say a quick hello before we move on. Hello, everyone. I’m James. I’m the global product support manager here at Opus two. I’ve been here for about a decade now delivering support to our end users. As we’ve scaled our AI offerings at OPUS two, more and more of the support workflow that we’re seeing is dedicated to assisting end users with AI query and how to get those best results, so becoming more and more common. We always suggest the Tidec format, something we’ll go through later on in these slides. Perfect. Thanks, James. So yeah. So I think from my perspective, I bring a legal and governance lens along with a sort of practical perspective of using AI for legal work day to day. And James, as he just said right now, works directly with our clients, supporting them on, best practice when they’re deploying and using AI across their organizations. So between us, we plan to cover how the how to use these tools responsibly and how to get strong results from them. Before we move on, just a quick one. We have about like I said, we’ve got about an hour today. It’s quite a lot of content. Please do ask questions in the chat. If we’re not able to get to it in in kind of during the session itself, we will absolutely come back to you afterward. So please do ask away. So, as this slide says, AI tools respond to what you ask for. And I think lawyers start ahead here because it’s drawing on what we already do. So applying logic, attention to detail, clear instructions. You know, none of this is new or novel. It’s building on what, you know, skills that we use and we’ve learned and we deploy it every single day. Now I don’t think we could have a session on AI without talking about hallucinations. The concern about hallucinations is real. You know, we have all seen the headlines, fabricated cases, hallucinated citations. But I think the same discipline that you apply and we apply to our legal work is what can help keep the AI in check. And I really think about it in two ways. One is how you structure the ask. We’re gonna cover that a little bit later, and James is going to deep dive into some practical example of, examples of what that looks like in practice. The other part is, and this is the key part, verifying what comes back against the source and that being verified by a human and not by an AI tool. And I think if you do those two things well, what ends up happening is you stop hoping for the AI to be right. You start managing it in the way that you would manage any other work product. So before we move on, I think James is gonna run a quick poll to get a bit of a read of the room. I’m gonna move the slide ahead. One second. There we go. Over to you, James. Thank you. So I can show your answers in the chat. There’s no wrong answer here. But, effectively, how confident are you in your ability to write effective prompts for legal work? Give a couple of moments. I can see the chat. Okay. Awesome. I’m seeing a lot of threes and twos. Awesome. So I think, you know, as a growing trend in the legal industry, what we’re seeing is more and more people are using AI daily. And as a result, they’re getting more and more confident with using the tool. The issue is then how to get those consistent results with each prompt that is used. That starts with some of the duties the attorneys already work under. So I’m gonna hand that back to to go through. Thanks, James. I think the other thing I would say, I mean, before we get onto this is and and seeing the kind of twos and threes in there, It’s, it’s very easy when you sort of see everybody talking about AI and how much they’re using it to to think that sort of everyone’s over here and they’re, you know, building agents and, you know, the rest. I think it’s absolutely fine to sort of start where you’re at, start small, and then build from there, and everybody’s doing the same. So, like I said, on the previous slide, two two parts of sort of responsible prompting. One is setting a structure, and two is verifying the output. And and the key principle here is that the lawyer stays accountable. So all of the practical techniques and tips that we’re going to talk through later on in this webinar all sit inside the legal duties that you already owe. So competent, can candor to the tribunal, confidentiality, and supervision. None of that changes because the work was drafted by AI. And the thread running through them is all the same. You’re gonna hear me say this quite a few times on this webinar. It’s a human in the loop. The model can draft, and it can surface options, but it what it can’t do is that it can’t exercise judgment. That has to stay with you, and that responsibility does stay with you. Keeping a human in loop as well is is not an abstract. In practice, it just means we’re checking the tool’s work so we can make that concrete with each technique that follows. Yeah. Absolutely. And so the sort of legal principles and and and legal basis around it. So prompting carefully and verifying the output, I think everybody should be doing that. That just gets to better results. But for lawyers, they’re how you meet your, legal duties as well. And, you know, most, if not all of you, will be familiar with the formal opinion published in twenty twenty four, which was the first formal guidance issued on generative AI. And interestingly, formal opinion five one two did not create a separate AI rule book. It took the existing model rules that, you know, that were already in place and applied them to AI. So nothing new, you know, that no no new rule book created. So if we dig deep into one of them, so model rule one point one, a lawyer shall provide competent representation to the client. We know this. You know? Like I said, it’s rule it’s rule one point one. It’s one zero one. That that’s what lawyers do. But technological competence is not new either. And the ABA made this explicit sort of more than a decade ago, in in one of its comments and said that, competence applies to the technologies that you use in, in your practice. So the duty already exists. It’s now just being applied to AI. The good news is that competence in the context of using AI doesn’t mean understanding all of the intricate mathematical details behind a model, and you don’t need to know how to build one either. Competence means knowing what the AI tool does, knowing what its limitations are, so where is it going to fail, and knowing how to check the work. And the sort of practical consequence of this is in the context of this rule, poor prompting sort of stops being just a quality issue, and it becomes a competent it becomes a competency issue the moment that work is going out of the door unchecked. You know, you wouldn’t do that with any other work product. We shouldn’t be doing it with AI generated work product either. We’re gonna spend some time in the second part of this session covering that exactly. So how to prompt within those limits, and just as importantly, when and how to verify the information that you get coming back. Yeah. James has got some really good, insight into sort of how to put that into practice as well. So rule three point three, candor towards the tribunal. Really simple. What you put before the court must be true, and it must be verified. Like I said, we’ve all heard about, hallucinated citations, fabricated cases. But when a model, it kind of hallucinates a case, it’s not lying. It doesn’t know the concept of lying. It is basically just creating an output based on the text that it holds. And, you know, one thing to remember is AI is absolutely brilliant at being confidently wrong. The models are getting better. They’re definitely getting more controlled, but AI is very, very good at being confidently wrong. But, ultimately, that, you know, that that’s not really helpful for lawyers because, like we said before, that responsibility stays with you. You know, this is where all of the sanction stories and, you know, headlines come from. And I think every single one of those was avoidable had the outputs been independently verified. So we’re gonna move on to confidentiality. And this is one of the, you know, rules that firms and lawyers worry most about. You know, the duty of confidentiality doesn’t just pause or stop because you’re using AI. You know, it’s not just about pasting it into a chatbot and then, you know, what happens next. The obligation is the same. Right? Whether the information is sitting in an email, sitting in a memo, or you’re putting something into AI. And and that can be the risk with free and consumer tools. You know, the ones that you sign up to in a minute, where you don’t control or know actually what happens to your data once you’ve submitted it. They can be quick to kind of get up to speed on and and sign up, but it’s sort of a bit of a vacuum, once you’ve actually submitted your data. So I think the first question I mean, at least for me, the first question I I ask isn’t, is this tool useful? The first question I ask is what’s going to happen to my data? How is it going to stay protected? So, yeah, I I sort of recommend that you you do the same. And this is sort of part that causes a lot of confusion. When you’re looking at a free or an enterprise kind of legal tool, both of them look identical out of the box. But underneath, that’s where the the tech can really change and things can matter. So on a free chatbot, for example, how long your data is retained, whether it’s used to train the model, who can access that data can vary quite significantly from an enterprise level legal tool. So the the rule of thumb really is gonna be simple. You need to make sure you know the tool’s data terms before you add any kind confidential material to it, and then treat the free tier as a kind of riskiest option until you have proven otherwise. Yeah. Absolutely. And I think just building on what James said, think about the data that you’re going to be submitting, you know, case strategy, client documents, privileged material. That that’s really in essence what rule points, one point six is seeking to protect. And you’ll see on the slide there, you know, outside of confidentiality risks, there are real privilege risks as well. You know, putting privileged documents into uncontrolled systems risk ultimately risks waiving privilege. We’ve seen that happen in actual cases. And it’s not a risk that you want to discover after the fact. You wanna make sure that what you’re using, you know what what’s happening to your data, how it’s being protected, before you put anything in there. In practice, this is gonna come down to really one area you wanna check. Before you can submit anything, ask where the data is allowed to go. Approved tools typically in enterprise level exists specifically, so you don’t have to make that call every time you do a prompt. And like we said, that those free tools do have their place, but you will have to ask that question every time. And so it’s really not the best place for any sensitive work. Yeah. Completely agree. So the lawyer stays accountable. The data’s you know, stays where it should be, stays where it’s protected, which raises the next question. So if the lawyer’s accountable, you know, there were there were, four competencies that I referenced at the start. Who is supervising the tool? So that’s where the next set of rules come in. There are two rules here. Rule five point one covers supervising lawyers, and rule five point three covers nonlawyer assistance. So that now includes the tools and technologies that you use. And prompting itself is a supervisory act. You’re effectively delegating a task to a system to complete. When you’re extracting a model, you’re doing something every single lawyer does regardless of level. You define a task, you set the limits, you explain the output that you would require, and you review the result before it goes out. So whether or not you’ve ever been supervised with by somebody or not, you already have a good idea about what good, clear instructions look like because you’ve either been receiving them and something’s not been clear in your past, or you’ve been, giving them, and you wanna make sure that your instructions are as clear as possible. So, again, I come back to you. This is not a new skill. This these are skills that you already have and that you deploy in your day to day already. Exactly. When you look at this When you look at this slide here, we’ve got some three key points, direct, constraint, and check. And just as if you’re providing conversation to someone in your team, you wanna with the direct, you wanna make sure you tell them precisely what you want. So the scope, the format, the task itself, vague instructions are gonna give you vague results. Constrain means setting the limits. So what to use, what to avoid, what sources to rely on, etcetera. You do that with anyone in your team. You’d give them areas to work within so that you’re limiting the scope of of what you’re getting back. And then check is just, again, that that kind of keyword verification is that you’re verifying what comes back before you rely on it. As with any work that you don’t produce yourself and even work that you do produce yourself, you always want to verify. So this step is not an optional one. It’s it’s always one that you have to carry out. So I really kind of rounding that put simply. A good prompt is just a good set of instructions similar to if you were delegating work internally. You’re gonna provide that clear scope, some clear constraints, and and then a a good way to verify the output at the end. Absolutely. So the the final rule that we’re going to touch on is one point four, so communication with clients. There’s no single answer on this. It depends on the scope of the work, your engagement terms, and what clients expect. But I think the the direction of travel I very much see is absolute transparency on where AI AI is being used. Yeah. That that’s certainly the direction of travel that we’re seeing. Yeah. And I think being able to answer clearly to your clients is is really gonna be the the trust builder when it comes to where the information is going and how the information is being generated. It’s it’s not a concession, so to speak. But hearing that, you know, you use these tools deliberately with the verification, with the oversight is a kind of it’s a reassurance rather than an alarm because when it’s a surprise, that’s gonna be what damages the trust of the client, not using the tool itself. Yeah. Awesome. So that’s the sort of legal framing. I think if you take one, you know, one thing away from this part of the webinar is the same duties that you already carry apply to your use of AI. Confidentiality, candor to the tribunal, competence, and supervision. So, again, nothing new, nothing novel. I’m gonna hand over to James now, who’s gonna demonstrate how some of this, works in practice. Thanks, James. Awesome. Can we get to the next slide, please? Okay. So during the next portion, what we’re gonna do is we’re gonna cover why attorneys already have the raw skills required for good prompting. So common pitfalls and how to avoid a repeatable how to avoid those and the repeatable framework as well as a few techniques to reduce errors and hallucinations. Oh, next slide, please. Thank you. Before we conclude, I’ll also show a brief demo in OPUS two cases to illustrate what good prompting looks like in practice and how that can be applied to a workflow. Before the tips, we do wanna hear from you again on your prompting challenges. So what are you finding frustrating about prompting right now? If you wanna check that in the chat, we can have a quick read through some of those. Yeah. Yeah. For one. Yeah. That’s a common one, isn’t it, James, about phrasing follow ups? Yes. So I think we’re gonna address some of these in the next few slides. But looking at a lot of these answers as the consistent format and the output, the multitude of questions, then also getting the correct answer with follow-up prompts seems to be a common one as well, And hallucinations. Yeah. That’s another big one there. Okay. So I guess we’ll cover these in these slides, and, we can we can address some of these these concerns as well, some some tips on how to avoid those. Do want me to move you along, James? Yes, please. So prompting versus prompt engineering. Prompting is very much just typing a question and seeing what comes back. Prompt engineering is about designing, testing, and reusing the prompt time and time again. A short comparison really between the two is a one off prompt is gonna produce one off quality, while reusable prompts are gonna produce more consistent quality each time. Legal work is full of recurring tasks, deposition summaries, chronologies, etcetera. And so what you really wanna do is build that prompt out once, refine it, and you’ll get a consistent output instead of rolling the dice for each matter. Yeah. And I think, you know, I I think about it also as a difference between sort of even, like, a one off email or a one off, you know, a contract that you draft once and then start from scratch every time versus building a precedent that you can use over and over again. So as we said at the start, there is no single best way to prompt. But like I said again, many of the things you do day to day, are already what you need for good prompting. So precise language, you draft to remove ambiguity because you know that imprecise language can create risk. A prompt works the same way. You know, the clearer the instruction, the more relevant and accurate and effective the output. Anticipating misreading, this is, you know, this is a really good one. So you draft a clause, now against how the other side might view it. You know, with a prompt, you you think about how the model is going to react to that and and think about how the model might interpret it in the wrong way and and catch it before before you submit. Spotting issues. So, you already verify your sources rather than take them at face value. That’s exactly the same discipline, that you can use to your use of AI. You know, like we said, hallucinations, duty of candor. Sequencing. So I think somebody said it in the chat. So the like, getting a response that’s on target, you know, sometimes that can be difficult if you’re you’re processing across a large volume of documents or your task is quite complex. So breaking those tasks down into smaller pieces reduces the scope for error and allows you to sort of build, from your initial starting point. And I think in our experience, lawyers that struggle with prompting are not lacking the skills. They just haven’t connected the skills that they already have to their use of AI. Yeah. And we we see this in in practice. The strongest prompters that we’re providing support to are usually the most careful when it comes to building out that prompt. There’s not really a case of people trying to chase any sort of clever trick. It’s just being precise and anticipating what the AI model is going to do with that prompt and where to check that the results you’re getting are verifiable and and correct. Yeah. And and that’s sort of the more you do it, the more the same things keep coming up, and it’s just it’s just that repetition. So as James said, there’s a few common pitfalls when it comes to prompting. So I’m gonna move on and let James cover those next if my slide will let me move on. There we go. So here’s an example of a prompt to summarize this deposition and tell me what’s important. It feels reasonable, but there’s lots of questions that could be pulled from this. It’s important to who, what issues, what should it cite, all of those things. So, actually, when you would provide this to an AI model, it’s it’s gonna be inconsistent results each time because we’re not providing it that additional information with the initial prompt. So many prompt people are writing prompts ad hoc, writing questions in the moment, seeing what comes back, and there’s no real structure repeatability there. The model itself won’t push back. It will just fill the gaps. So that’s where your errors, your hallucinations, and such are gonna come from. In this case, you know, it might decide that a particular issue is important and only focus on that issue throughout the entire deposition and ignore everything else. So what you really wanna do is kind of treat the the model as a new hire. So someone that has no memory of your norms when it comes to processing documents or reading through depositions. It only knows what you provided, so the briefing really has to be explicit here. So what you really wanna do is make sure you include things like clear structure, a clear framework as well, which we’ll cover shortly. And then as a result, you’ll get reusable prompts that provide consistent and reliable output across such recurring tasks. Yeah. Absolutely. And, I think James said it earlier on this slide. Just think about how you would delegate to another colleague and apply that same level of precision when you’re using the AI tool. So there’s also prompting frameworks that you can use. If you go into Google those, you’ll you’ll see lots of different options that depend on the models you’re using, depending on what sort of outcome you’re looking to get. We suggest the Tidec framework as this one that’s well suited to quite a few legal tasks. There are six elements to this. There’s the task type, the instructions, the do’s and don’ts, so the positive and negative constraints. Also, section four providing examples in your prompt of what you’re looking for, and then the content itself. And depending on the tool you’re using, this might just be provided as a file or it might be something that you’re copying and pasting into the prompt itself. Build it building these out is great. Once you’ve built them out, you’ve tested them, you can save those for recurring tasks, and then you’ve got a reusable prompt library at that point. So this is most useful for those repetitive tasks that you’re doing day in and day out. Absolutely. Yeah. And and, you know, again, almost think about it as a delegation checklist. You know, have you covered and ticked off all of those items? And, again, just going back to some of the comments around the sort of prompting challenges, I think, James, this is exactly the kind of framework that a lot of those that would address a lot of the challenges that have been posted. Yeah. Exactly. So breaking them down a little bit more here. Task type, what does that mean? Task type is gonna set the operating mode up front for the AI to perform. So what kind of job you want it to do. Do you want it to summarize? Do you want it to extract? Do you want it to draft? And you’ll provide that clearly as the task type within the prompt. You don’t need to create a flowing paragraph or essay to the the AI model. You can simply put task type and then the task. And we suggest that. And when we get on to an example, you’ll see I’ve actually just kept each of these framework points in my actual prompt itself. The AI model doesn’t see those. Instructions, so they define the expected action and scope and and the goal. This is the main body of the prompt itself where you’re gonna define the deliverable that the that the prompt is gonna output. So for legal work, you could do some additional parts here. So naming the jurisdiction or the governing standard in the instruction. This will allow your analysis to be a bit more anchored rather than generic. And then if you were to take these sort of task type and instructions that we’ve done here and see that in comparison to a summarize this deposition prompt, you’ll get a considerably different output. You can then steer that output with the dos and the don’ts. Yeah. If we get an next And I think I think the one one of the things that I found incredibly useful is you can see in the task type, you know, actually just writing your persona. So who are you? And that helps the AI assume your sort of profile attorney reviewing a contract. And it’s just one more way of helping make that output a little bit more effective. Okay. So do’s. These are gonna shape your output. Like I said earlier, positive constraints that you want, and this is gonna be your styling, your structure, your tone. And the example here, we’ve said for each argument, give the defendant’s position in one or two sentences, then our likely counter, then a confidence label of strong, moderate, or weak. So this is how it’s gonna break down the response that it gives to you and gives explicit instructions on how it should structure that response. Don’t so the negative constraints. This is gonna prevent failures effectively, what you guys have said in the chat, hallucinations, complete misdirects, or going well outside the the point that you’re looking to achieve. So in here, you can say, you know, use only the documents provided. Don’t rely on outside knowledge. Don’t don’t pass over citations. Don’t invent facts. And this is where you can also put in information like flag where a fact or authority appears to be missing rather than just filling in the gap. So where there might be data that’s missing or or we’re looking for the AI tool to let us know that that is happening rather than just hallucinate and and fill in that gap. Yeah. Absolutely. And, again, if we if we think back to the principles that we spoke about earlier, the the dos and the don’t, particularly the don’ts, are helping doing that professional responsibility work, setting the guardrails, reducing the likelihood of out of scope inaccurate outputs. Move us on just how are doing for time? Good. So examples, these give the models a pattern to match. So even a one liner can show the output format that you’re looking for, and that’s gonna lift the quality significantly. They don’t need to be lengthy formats. Even those short examples will do enough. And then the content is simply the raw material that you’re looking for the model to look at. So whether it’s a transcript or a document that you attach or just text that you paste into the prompt itself. In addition to this, there’s a few other techniques that you can apply to take this a little bit further as well. Is my screen still sharing? No. I think it’s it’s lost on that one. Oh, sorry, guys. Sorry. Where where is presenter view? Slideshow. Slideshow. There we go. It is slow. James, we’re gonna try and do the next bit in about five minutes so we give you time proper time for the walk through Okay. From current slide. There we go. Sorry, everybody. No. You’re good. Go one more side. There we So just building on that, I think a couple of people have have touched on in chat where they’re trying to do larger tasks and getting difficult responses there. What we’d really suggest is avoid trying to do everything with one prompt, taking a huge document set and throwing a large prompt to that five, six, seven plus from the queries in that one prompt, you’re gonna get a significant drop in quality. So what you want to do is break it down into steps and verify between each step. So when you’re going into each step, you verify the data, which reduces the risk for a hallucination passing all the way through to the final task. A useful prompt breakdown, say, for example, a chronology, a second then checks each entry against documents, flags, and issues, and a a third prompt then fixes what’s been flagged. It’s a small, repeatable, and checkable steps instead of one big ask. Absolutely. And, again, going back to the principles, we’re linking back to supervision here. You’re you’re breaking the task into small, manageable pieces versus one big complex exercise. Let’s go quickly. We’re sharing the slides, aren’t we, afterwards? Yes. So a chain of thought. This is a way to reduce hallucinations. You wanna ask the prompt your prompt, ask the model to reason before making any conclusions. This is gonna materially reduce the risk of incorrect or incomplete outputs because it’s gonna check itself effectively. The key instruction here is to ask the tool to list any passages it relies on, provide citations, etcetera, and to save when support is needed because information is missing. Yeah. And that sort of legal analysis, that’s effectively IRAC. So identify the issues, state the rules, apply the facts, and lead to a conclusion. So the third habit that we would suggest is making the model argue with itself. So if we get to the next slide. Once you’ve got analysis, make the model itself argue with it. So make the other side’s best argument and name the three weakest points in what you wrote. This is then building on the sort of prompt that you may have done in the initial test. This is a good way to quickly pressure test position before any opposing counsel does. Yeah. And I think watch for incorrect assumptions throughout. So the model may accept a premise in one of my like, one of your earlier prompts and then build on it. You know? I’ve I still have it today. You sort of you’re in a kind of conversation that, you know, send me messages down and you spot something and you realize that the model created an error right at the beginning and sort of just seen that error through. So being able to look out for those is really helpful. And then grounding. So building verification and checking into the prompt setup, which is both, you know, good practice, but also really effective because you’re not waiting to the end to check. You’re building in that verification right from the beginning. So answer using only the documents provided really, you know, feels like an obvious one, but it’s really, you know, really, really useful. Do not use outside knowledge. Again, only review based on the documents, that I have provided to you, and don’t make assumptions. And then be precise in your question. I mean, James and I have said this a few times. So it’s don’t, you know, ask whether these documents address x versus why did something happen. The the difference in the first one is it’s much more of a specific question, and the AI is going to be much more effective at processing and pulling the information and giving you a much, much more accurate output. And that’s gonna bring us into hallucinations and assumptions and how to look out for them. So a hallucination, this is when the model is gonna give a confident but usually wrong answer with incorrect citation or quote. These are failures with real consequences. The techniques we’ve covered really cut the rate down of hallucinations. There’s that tight scope, the grounding, the stage prompts for each part, and then the verification between their steps. Yeah. And, you know, Florida this month and New York this month just made this explicit. So in all filings, every filer is required to certify that the cited authorities exist and that they’re accurately cited and that they have been verified. So, again, being built in to the regulations sort of in in real time. It’s replacing a patchwork of local orders. You know, more and more states are sort of issuing their own legal rules on how AI should be used, and almost all of them so far center around verification. So it’s just something to keep in mind. And then finally, we’re gonna just get an example of how all this comes together in a prompt that could be built into your workflows. So repeatable is gonna mean staged, but not one giant asks. You’re gonna go from pleadings to issues, building a chronology, strengths, weaknesses, and stress test. Each point here, you’re gonna verify and make sure that the information you’re taking from one stage to the next stage is correct and, overall, reduce any hallucination risk and how you then satisfy Candle. Yeah. And, you know, so far, we’ve spoken a lot about kind of what the AI is doing. This is absolutely assisting your strategy, but it’s not setting your strategy. Right? That that judgment, that ownership stays with you. So let’s jump over and demonstrate a prompt. I’m gonna stop sharing my screen. I hope I don’t break this. No problem. So I’m gonna demo this in Opus two. For those unaware, Opus two is an AI enabled legal case strategy and collaboration platform that helps dispute teams organize case materials, develop strategy, streamline workflows, and deliver those outcomes across the dispute life cycle. We’re gonna use focus view, which is a streamlined AI first UI to interrogate case documents, synthesize information, and draft outputs within Opus two. The example I’m gonna use is what we spoke about previously. Me share my screen. So previously, we spoke about building a prompt for asking to summarize a deposition and tell me what’s important. So I’ve actually built out a prompt for that using our framework, and I’ve saved that here. I can talk you quickly through it. In here, I’ve defined my task type. So this is a deposition transcript summary and litigation support issue spotting. Instructions. And here, I provide instructions on what to do and what to look for in terms of indicators within the provided files, define what’s important within that transcript, what I’m looking for. And then I’ve got my positive constraints, source checking in here, as well as my notes in here as well. Here we go. So in here, for example, do not provide legal advice. Do not invent facts. People do not assume the jurisdiction, etcetera. These are all very clear explicit rules of what we don’t want the AI to do. And then I’ve given some examples here of what a strong and important point would look like. So the witness confirmed x, an example of a weak point or the witness acted negligently. And then I’ve given some examples of what a better version would look like for those. And then I’ve got my user content, which is here. I’m just outlining to use the selected transcript. And then if I wanted to, I can actually edit this prompt to include some key other information that I may have, which is just gonna provide further context to the model. So, for example, if you have a list of issues already, you could use those here, and you could provide those and ask it to look for that. So paste it in here. I’m just gonna select a transcript. I’m gonna run that print. You can see I’ve just kept in my headers. Like I said, you it doesn’t matter to the model. You can just include those, and then make it easier reading for you. And I think would it be fair, James, to say, you know, it it’s it’s really down to the user as to how much detail they want to put in. So there, that’s the sort of headlines under the framework, but you can sort of start smaller and then build from there. Yeah. In this example, I went really heavy just to do an heavy example. Yeah. But in in I think I replied to someone in chat for this. You could simply do a sentence per point in the the the Tardot framework. It doesn’t have to be as detailed as this. But the more detailed you are, the more detailed response you’ll get. And you can see here, it’s giving me a lot of information, and it’s really going into depth based on the examples I provided, what I want it to look like, and I will get repeatable answers for this because of that framework and because of the boundaries that I set for the model itself. So don’t feel that you have to do paragraphs upon paragraphs for every single type of query. This is very much just an example to show how aggressive you can be. Yeah. And I just I did see in the can you expand the text of the prompt? Oh, I’ll make sure that the prompt is sent over as well with the slides so people can have access to that. Cool. Yeah. That’s a it’s a I mean, it’s a lot of info, but it’s also I mean, seeing it really well structured. Yes. And easy to walk through as well, alright, versus just a block of text that you might otherwise be navigating. And I think that’s something to also kinda mention with your instructions. I’ve asked it to provide it in table format, for example. Yeah. If you don’t ask it to provide it in that, you’re just gonna get bullet points. We were just gonna get text. And you can be very explicit with most tools at this point in terms of how they format their output. So if you wanted it to say something like in markdown format, so you can copy and paste it into Word, for example, or in a table format so you can paste it into Excel. You would just include that in your do’s or don’ts or your instructions, and it will make sure that it does that for you. I think we’re getting some kind of end of my prompt at this point. And in here, you’ll see the support check it’s doing on verification and hallucinations. That’s awesome. I mean, the one thing I mean, a lot of, you know, many enter legal specific tools do this. But if you scroll up, James, the the interesting part also is the source is clearly referenced. So as you’re going through the document, that that step about verification, you can simply click in. You know, you don’t have to go back through to troll through where that came from, which is Yeah. And in this case as well, I did ask her to look at any kind of exhibits Yeah. That might support what’s being said. You can see that’s done that here, and it’s provided links to those sources as well. It can be downloaded and used offline. Okay. So now that we’ve covered the rules, the best practices, and I’ve given you an example here of how this could be applied in a case strategy workflow, we can summarize with a few key takeaways. Perfect. And I think probably just before we move on, the only other thing to mention is once this is created, you don’t lose it. This pushes back into the platform. Right? And it’s there for you to use, go back and review search against on a sort of ongoing basis. Yeah. Yeah. I have this available here, and I can save this conversation Yeah. And keep it in the platform. And my prompt here is also saved, ready to go next time I wanna use it, whether it’s for a different document or even the same document again. Say I replace it with an updated transcript in the future. And I guess, James, I mean, we’ve looked at kind of a prompt in isolation. Right? How would this fit into the sort of bigger case case strategy workflow? Where where would this sit? Yeah. A good example is, say, chronology building. What you may want to do in those cases is instead of just saying build me a chronology, if you have maybe a particular subset of people that you want a chronology based around or particular company or whatever it might be, you could build out these models to different stages. So you could do an initial prompt to provide instruction to fetch all of the sources that relate to a particular time frame or particular group people or particular company. And then once you have those, you can then do your secondary prompt on building a chronology from that information. And then at the end of that, you can do the fact check on that final chronology where it will refer back to the original documents provided and ensure that nothing has been changed in those three separate stages. Got it. Awesome. Perfect. James, while I share my screen back just to do the key takeaways, it might be worth just touching upon how this actually feeds back into the chronology in the in the main platform? Again, so you’re sort of not having to redo the work? Correct. Yeah. So from here in the platform, I can actually go straight into individual documents, and I can actually then add them to a chronology, or I can download the documents to review offline. As well as if I want to, I can ask for for example, if I asked in my original instruction for some timeline here, I can take this data, and I can import it into the chronology with my source information and just link that automatically rather than having to rebuild the chronology itself. I can just use this. Awesome. Alright. Let’s have another go at sharing my screen. Do you want to stop sharing, James? There you go. There we go. There we go. Can you see that? It’s coming up? Yes. Cool. So, I mean, I hope that I hope you all found that interesting and then bit of a sort of prompting in practice. And I think as as James said, there is no kind of perfect way to start. You can start small and build build from there. But the more detail that you give, the more accurate, the results will be. So I just as we finish up, I just wanna bring this back to where we started and and and sort of close us out. I spent the, you know, first part of the webinar talking about the legal duties. The legal duties have not changed. Competence, candor, supervision, and confidentiality, communication, that all the obligations that as lawyers, apply to us today regardless of whether we’re using AI. Competence using, knowing what the tool does well, but equally knowing, where the tool has limitations, and how to check those limitations. Candor, verifying that what you’re submitting is correct, checking the work product before it goes out the door just as you would with any other work product. Confidentiality, a really big one. So, you know, not just submitting data into a tool, understanding where that data is going, how it’s going to be used, how it’s going to be protected. Supervision, you know, directing, constraining, and checking the tool just as you would, delegate any other task. Think about that when you’re using AI as well. So I think, you know, in in my view, the rules aren’t a blocker to using AI. They actually help you use them effectively and responsibly. You know, I’ve I’ve as James has been giving the practical examples, I’ve been tying it back to the rules. So, you know, it’s a it’s a nice it’s a nice set of guardrails, as you use AI and sort of feed to get more confident and and feel more responsible as you’re using it. So that’s sort of from my side. James, what would be your takeaways? Am I moving along? Let’s go. So for me, you know, when I’ve been working with people in support, the main kind of area that I’m helping with is just that prompt engineering that I spoke about earlier. And using something like the the Tidec framework, you’re gonna reduce the margin of error and give yourself that more consistent output. That being said, obviously, like I said, go Google different frameworks. You might find one that works slightly better for for you and the way you work. But using a framework is gonna be key for that consistency and that verification that you’re looking for. And finally, like that that, again, that key point, going one step further and taking the time to build out prompts for each stage of the legal work rather than just jumping to the end stage, again, is gonna help reduce those hallucinations and the assumptions and and make verification a whole lot easier for you. So taking that verified data into each stage with each prompt is gonna leave less room for error. So bare minimum, if you don’t build out prompts per stage, verification is gonna be key. Really? I think we’ve got one more slide before we close out. I guess, you know, while the fear around hallucinations and fabricated cases, may be real, hopefully, what you sort of taken away from taken away from today is that you can, as you you are using AI, build those checks, build those balances, and apply, you know, apply the same discipline that you would to any other work product, and it it helps keep your, use of AI responsible. You know, applying structure, verification, and a human always checking the output. I think James and I have probably said between us about ten times on this webinar, but we can’t sort of label that enough. The human in the loop is absolutely critical. So thank you, everybody. I know it was a packed session. I’ve been following the chat as I’ve been able to. If we weren’t able to get to all of your questions, we absolutely will, respond after the event. I think what I would say is I’ve learned a lot, about AI from sort of reading and from articles, but what I’ve you know, where I’ve learned the most is from speaking to other people, you know, understanding how they use it, what works for them, what doesn’t. So if any of this has sparked an interest or you have any questions, then, obviously, please do, connect on LinkedIn and be very, very happy to compare notes and answer questions. James, I’ll hand over to you. Yeah. Feel free to reach out to me on LinkedIn as well. Happy to kinda give some more examples around the type of prompts you can build as we’ve we’ve done a few of them in in support as as as examples over the past few months. But, yeah, thank you for joining. I hope you found it valuable in in applying new skills for stronger prompting, and it gives you the confidence to avoid any common pitfalls. Thanks again to ASEDS for hosting us for this session. If you have any questions regarding the CLE credit, I think you can contact ASEDS at customer support at ASEDs dot org, and have a great rest of your day. Awesome. And I think we’re gonna hand back to Mike. Yes. Thank you. Thank you, everyone, for joining us on the ASEDs webinar channel today. Thank you, James. Thank you, Smriti. Fantastic presentation. Everyone, you can learn more about, opus two at opus two dot com. Please visit there, to learn more. And please visit aceeds dot org for a complete list of upcoming webinars. Have a great day everyone, and as always, be kind to one another.