Transcription
Most lawyers are using Claude like a glorified Google search. They type a question, they get a wall of text, and they wonder why AI feels useless. That's not a Claude problem. Typically, that is a prompting problem. And here's what most firms don't know. Claude has five features built specifically for legal work. And I'm going to bet that your firm is maybe using one of them, maybe two at most.
I'm Liam Barnes, founder of Cyberactive. And for the past four years, we've been working with SAS businesses and legal teams, and helping them identify the parts of their operations that can be automated where they can effectively leverage AI. I've trained legal teams on these exact features. I've watched partners go from "AI is a toy" to running entire due diligence workflows through Claude in a single session. The difference isn't the tool, it's knowing which features and how to use it.
Today, I'm going to walk you through five Claude features that actually matter for legal work. Not a product demo, a field guide. I'll show you exactly where to find each one, how to set it up, and what it looks like with real legal documents. So, let's get into it.
Feature one is the foundation that really everything else builds on. And if you get this right, the rest of the video is going to be 10 times more valuable to you. It's a five-part prompting formula. You've got context, task, standards, tone, and format. Let me show you the difference with a working example.
So, you'll see here on the left side of the screen, I've just got a simple prompt, and this is typical of what most people do: "summarize this contract." And Claude gives you a wall of text. Technically, it's accurate, I guess, but technically useless. And I must preface this by saying I'm not a lawyer, but having worked with lawyers, I certainly got a good sense for what they deem to be good quality output and what they deem to be largely useless or of not much value.
Now, watch what happens when you give Claude structure. So, on the right side of the screen, I've got a different session of Claude open, and I've uploaded the same contract. It's a commercial lease agreement. And instead of saying "summarize this contract," I've said: "You are reviewing a commercial lease for a midsized retail tenant. Identify all clauses that deviate from standard tenant favorable positions. Flag anything that exceeds market norms for retail leases in California. Write this for a senior partner who will present findings to the client. Format as a table: clause, opposition, market norm, risk level (either red, yellow, or green)." And so I've read through this and it's certainly more detailed, but not only that, you'll see that Claude has generated this output for me, which is the table that I asked for. And so I'll just open that now. So you can see here we've got a full executive summary with a table with our position, market norms, and the risk categorization. And that's using the exact same tool, but just with a different prompt, with a better quality prompt.
Context sets the scenario, task defines what you actually want, standards gives Claude a measuring stick, tone controls who it's writing for, and format tells it how to present what you want or how to present the output. So you learn this once, you use it every single day. And it works on every AI platform, not just Claude.
Now let's take a look at how you can do this without having to write it out every single time. Feature two is projects. Think of projects as one workspace per client matter. And to access it, you just expand the little side navigation menu here. I'm in Claude; in other words, the web application, but it's the same feature. You access it the same way using the desktop app. So, I'm just going to click into projects.
And here's the problem that projects solve. Every time you start a new conversation with Claude, it forgets everything. You upload your documents, explain the deal, set the context, and then you do it all again tomorrow. Projects essentially fix that. So, we're just going to come over here to click project, new project. And I'm going to create a project for a fictional M&A deal, Acme Corp Acquisition. So, what we're going to do is just name the project and I'm just going to save the project.
All right. So, we're inside the project that I've just set up. Now, I set the custom instructions and these apply to every single conversation inside of this particular project, which means I don't have to repeat myself. And so, what I'm going to do is just give a little bit of information. "Client is a buyer. Purchase price $45 million. Flag anything affecting purchase price or earnout structure. Apply California law. All documents dated after January 2025." Then I'm going to save those instructions.
Next, I'm going to upload deal documents: purchase agreement, LOI, due diligence checklist. These are now permanently available inside this project. So, I'm just going to click that and I'm going to grab the documents and the stock purchase agreement.
Okay, so I've added the custom instructions. I've added the deal documents. Now, I'm going to write: "What are the key risk factors in the purchase agreement?" Click send. Now what Claude should do and what it will do is reference specific documents by name, apply the buyer's perspective from the instructions, and flag things affecting the purchase price or earnout. So you can see there we've got purchase price and payment structure, earnout risks, indemnification, representations and warranties gaps, structural California law issues, priority action items for buyer's counsel.
Now when I ask a question, "What are the key risk factors in this purchase agreement?," Claude already knows the deal. It knows I'm the buyer. It knows the purchase price. It references the actual documents I uploaded. And here's the thing though. If I just go back to the project. If your associates start a separate conversation inside the same project, they get the same context. Partners, associates, paralegals, anyone working off the same foundation. And you'll see all the conversations, each session will appear below. But as I said, every subsequent session has the same context and same information. Pro and max plans include projects. Team plans let you share them across the firm. Setup takes about two minutes as you saw right here.
Now, feature three is where it starts to get really interesting. We're going to talk about custom skills. And to do that, we're going to go over to the Claude desktop app. Now, a skill file is just a... a skill, I should say, is just a plain text file, a markdown document, just a type of document similar to a PDF or a Word document, just a type of document. And it tells Claude how your firm does things, your playbook, your standards, your red lines. And Claude applies them automatically.
So, what I've got here is a skill that I prepared earlier for an NDA review. It flags non-solicitation as red, California governing law required, cap term at 3 years, and a bunch of other stuff that feasibly could be specific to your firm or your legal team. And that's it. Plain English, no code, no developer needed, lawyer friendly. You do not need to be technical to understand and read a skill. Nor do you need to be technical to know how to develop and build an effective Claude skill.
In fact, we're going to create a skill right now. So, I've got a pre-prepared prompt right here. What we're going to write is: "Help me create a skill for reviewing engagement letters. Key checks: scope clarity, fee structure, conflict disclosure, retainer amount." Let's click enter. Or one thing I always advise lawyers who are starting to use Claude is to read the output so you can start to see what types of tasks and activities it's undertaking, what it's reading, and so you can just get a better sense for the research process or the steps it's taken to get to the output. And then after about 25-35 seconds, Claude's going to output the skill file. And so you can see here, if we just stay focused on the left hand side of the screen for a moment, the core skill with four review pillars and structured output: scope clarity, fee structure, conflict disclosure, and retainer, exactly what we asked.
So let's have a look at what Claude has output. So we can see here that... and so if we look at the skill, you can see it's really detailed, right? And as I said, it's just a markdown file, that's .md right there.
Now here's why I think skills are even more useful and more powerful than the likes of a custom GPT inside of ChatGPT, because they're locked inside of a web interface. You can't version control them. You can't share them as files. You can't put them in your firm's git repo. If that doesn't make sense, essentially you can't share them, right? So like code, skills are just a file. You can copy them, you can email them, you can put them on a SharePoint. A senior partner writes the NDA playbook once and every associate in the law firm can use it automatically. And so they're sharable and they're really malleable and essentially just a great way to standardize workflows and processes without needing to write code.
Okay, so let's do another example. So if you recall just a few moments ago, I showed you this NDA review skill, and I created that as a skill. And I'll just quickly show you one more time. So just come down here, go into your settings, go to capabilities, okay? And then you scroll down here to skills, right? And you just go add. Slightly different interface than the web interface, but not too different, right? And so just then we created one with Claude, right? We gave it a prompt and it created a skill. This time we actually uploaded a skill because we already had the skill in existence.
Now, how did I write this? Well, I didn't write it in totality. I wrote parts of it, but you don't even need to. You can just work with AI to write the skill for you in the same way that we just did there. But you want to get it into your system. So you can either—I'll just repeat myself so it's super clear—you can either have Claude create it and save it as a skill, or you can write up the skill document separately and then upload that as the skill. And so that's what we've done here, right? And so our skill is right here. And so now we're going to put it to the test. Is this skill going to do what we want it to do? And you don't need to do anything special to invoke these skills. You just need to use natural language prompting. You just say, "Hey Claude, review this NDA," and it will naturally use the skills it has available. If you want to make sure it uses the skill, say "review this NDA and make sure you use the NDA review skill."
So, what we'll do is just quickly type "review this NDA" and we're going to upload an NDA. I'm just going to hit that button. We're going to add it from a local file and we've got a fictitious NDA right here for a company Pinnacle Systems Inc. Now, because we've got the skill, it's going to follow the instructions outlined in the skill. So, you see there it says "check NDA review skill." So, it's automatically invoked the skill and then it's followed the system outlined in the skill to review this NDA. And so, what you've got here is you've uploaded an NDA and Claude has applied your firm standards without being asked. It's caught the non-solicitation clause, flagged the 5-year term, and called out Texas governing law. All from a plain text file.
Okay. So, the fourth feature is one that typically gets the biggest reaction when I show it to partners, and it's called co-work. It's a recent release, and you got to—you can't be inside of a project to use it. You can see there's the three little indicators or three options at the top of the screen. So, you're just going to select co-work.
Most AI tools are back and forth, and you can probably relate to this. You ask it, it answers, you ask again. But co-work is different. You give Claude a complex job. It creates a plan. You approve the plan and then Claude goes and does the work. Creates files, builds spreadsheets, drafts memos. Let me show you.
I've uploaded three contracts and I'm asking Claude to compare them. I've got a pre-prepared prompt right here. "Compare these three vendor agreements. Create a spreadsheet with: vendor name, term length, monthly fee, annual cost, indemnification type (mutual or one-sided), indemnification cap, termination rights, SLA/uptime guarantee, data handling post-termination, IP ownership, governing law, and a risk flag (green, yellow, red) for each contract overall." Hit "Let's go."
Claude starts by showing me the plan. And what we're going to see here is Claude is going to show me their plan first. What they're going to do step by step. "I'll read all three agreements and the skill file. Now, let me set up the task list and build the comparison spreadsheet." So, it's a structured way of thinking that gets you better results for complex projects.
So, I approved Claude's plan. Claude goes to work. It read each contract, extracts the terms, builds a spreadsheet, and presents it for review. And so, what we've got here is here's the completed comparison. The risk flags at a glance: Gamma green, Alpha Cloud yellow, Beta Metrics red. Let's have a look at the spreadsheet.
Now, that's pretty impressive. I mean, that's probably a task that would take a junior associate, I don't know, half a day. Claude did that in under three or four minutes, and I got approval checkpoints along the way.
Now, to be completely direct and transparent, co-work is still a research preview. Anthropic says it's not production ready for regulated workflows yet. It's Mac OS only. You can't use it inside of projects and it's not captured in audit logs. But for internal work, due diligence prep, matter summaries, contract comparisons, it's really powerful and really, really useful.
Feature five is one built specifically for lawyers. It's the Anthropic legal plugin. I created a full video on this just last week. I'll leave a link to it in the description below. Essentially, what it does is give you slash commands designed for legal workflows. But here's what makes it different from generic AI tools: your firm's specific rules are baked in. So when you run these commands, Claude doesn't just flag problems, it flags things that violate your firm's specific positions.
And in case you didn't see the first video or you need a reminder, you can only access this currently on the desktop app. And so you're going to go up to co-work, okay? And then you go to plugins and then the legal plugins are—you just need to add them, right? So you come over here, browse plugins, and if you type in "legal," you'll see it here. And if you click into that, you'll see all the skills that are available, right? So you've got: review a contract against our playbook, triage NDAs, brief for an upcoming meeting, etc. Okay, so we're going to go out of that because we've already added that. Again, I'll leave that video in the description.
So now I'm going to show you how you can incorporate your firm specific positions with the benefits and the pre-existing skills that the Anthropic legal plugin gives you, right? So here's how you do it. So over here I've got for a fictitious firm legal standards. So we've got indemnification, governing law, non-compete, non-solicitation, IP ownership, etc. And so there's a couple of ways that you can do this.
So let me just explain again what we're going to do here. We're going to use the pre-existing skills that come with the Anthropic legal plugin. We're going to use the "review contract" skill, okay? But we're going to incorporate a firm's specific positions with the legal plugin. And so there's a couple of ways you can do this. And I'll try and make this sound as—or I'll try and explain this in as simple and easy to understand as possible. It's really basic, but there's a couple of ways.
So, first thing: what I could do is just paste this in, right? And then I'm going to upload a contract. So, I've uploaded the commercial lease agreement. And then I'm going to do "review contract" — slash review contract. You can see it there. I'm going to select it. And so, this is one of the skills that comes with the legal plugin from Anthropic. So it says "review contract against your organization's negotiation playbook. Flag deviations. Generate redlines. Provide business impact analysis." So when I do that, it's going to do everything that's in the skill, but it's also going to reference the firm positions.
Now, let me just show you the other way that you could set this up. So what you could do alternatively (and I actually think this is a better way to set up your Claude environment for ongoing use) is attach the agreement the same way that we just did. Type "review contract." So let me just upload that agreement again. Okay. Now instead of pasting the firm positions in the body, what you could do instead is connect a project. Now if you recall earlier we set up a project and you had—you can add files and folders and instructions inside of it. And so what you could do is add this as a file inside of the project. So you would select a project, add it like this, and then inside of that project you would just add that firm position document, right? So either way works, and I'd probably—well, I would suggest that this way is going to save you time over the long term because you're not having to copy your firm positions each time.
Okay, so that actually took maybe three, four minutes. But look at this. Every clause is flagged. Green, yellow, red. And here's the key. These aren't generic warnings. This red flag here says "one-sided indemnification." Why? Because my firm standard is mutual indemnification capped at one times contract value. The lease violates that. Red flag. Down here you can see "New York governing law." Red flag. Why? Because my firm requires California law. This lease violates that. Big red flag. This plugin knows your firm's playbook and enforces it on every contract automatically. I mean, this is really thorough, really detailed. And this is only possible—or this type of output is consistently possible—because you're combining pre-existing skills with your own internal positions.
And the legal plugin is great. The legal slash commands are really, really effective. My clients have really liked the output that it's been given on specific tasks. For example, the "triage NDA" gives you a quick approve or reject decision on an NDA. "Vendor check" lets you search for existing agreements with a company. "Brief daily" gives you a morning summary of everything legal relevant from your emails and your calendar.
The power here is simple. Your firm's rules applied automatically to every review. No typing them out, no forgetting them, no back and forth with an AI that doesn't know your standards. The plugin knows. And this launched, as I said, early February 2026, and it's on Mac only right now. And once you set it up in Claude desktop, it's there permanently.
Now, let's talk about what Claude can't do because I'd rather be upfront and direct with you. And you hear about this from me, and I give you my perspective, rather than find out the hard way—as in you get disappointed when you use Claude.
First, Claude still hallucinates. About 5 to 10% of legal citations it generates don't exist or are misattributed. That's better than ChatGPT's 15 to 20%, but it's not zero. You verify everything. That's a non-negotiable.
And second, there's no Westlaw, no LexisNexis, no paywalled databases. None of the AI tools have this. Claude works from what you upload and what it knows from training. You can of course build out your own custom automations and systems that tap into Westlaw and LexisNexis. And if you are interested in having one of those custom-built for your firm or your legal team, Cyberactive—reach out to us. That's exactly what we do.
And finally, the best features are Mac only right now. Co-work, legal plugin, file-based skills—all Mac OS. Windows support is expected mid 2026 though.
And one more really encouraging thing: Claude's context window just hit 1 million tokens. That's roughly 2500 pages of legal text in a single conversation. That's live right now as of February 18 on Opus and Sonnet 4.6. But that doesn't mean you should dump everything all at once. Structure still matters. The prompting formula still matters.
So let's recap. Five features:
1. The five-part prompting formula. Structure your prompts and the output transforms.
2. Projects. One workspace per matter. Context that persists.
3. Custom skills. Your firm's playbook applied automatically.
4. Co-work. Give Claude a job, approve the plan, get deliverables.
5. And finally, the legal plugin. Slash commands built for contract review with your firm standards baked in.
Claude Pro is $20 a month. Team plans are $30 per user. You don't need to roll this out firm-wide to test it. One person, one plan, one matter.
If you found this useful and you want to get more information about how to use AI effectively inside of law firms and SaaS businesses, I do a new video—sometimes two—every week. So, drop a comment if you have a question. I'll make sure I get back to you. And I'll see you in the next.