What Claude Skills and Cowork actually do, and why it matters if you're a lawyer

I've been testing AI tools for legal work for a while now. Most of the time, the experience follows a predictable arc: excitement, experimentation, mild disappointment. You paste a contract into ChatGPT, ask it to review against your playbook, and get back something that looks impressive at first glance, until you realise it missed half the issues and hallucinated a clause that wasn't there.

So when Anthropic released Claude Cowork and its Skills system, I approached it the way I approach most new tools: cautiously optimistic, fully prepared to be both overwhelmed and underwhelmed at the same time.

I wasn't.

First, the basics, in plain English

Before I get into what changed for me, let me explain what we're actually talking about. Because "Skills" and "Cowork" sound like marketing terms, as most AI product names are. But these describe something genuinely different from what you've been using.

Skills are instruction sets that give Claude specialised expertise. Think of them as training manuals that activate only when they're relevant. You don't need to upload them every conversation. You don't need to remind Claude of your preferences each time. They sit in the background and kick in when the task matches, whether that's reviewing a contract, building a spreadsheet, or formatting a Word document.

The clever part is that Claude doesn't load all of this upfront. It only pulls in the full instructions when a skill actually activates. That means you can have dozens of skills installed without clogging up the system's working memory.

Cowork is the mode that turns Claude from a chat assistant into something closer to a junior colleague. In regular chat, Claude responds to what you give it. In Cowork, you point it at a folder, describe what you need, and let it work through the problem independently. It reads your files, plans an approach, breaks the work into steps, and produces actual deliverables (Word documents, Excel spreadsheets, PDFs) saved directly to your folder.

That distinction matters more than it sounds.

What actually changed

In my previous workflow, I'd paste a contract into an AI chat, ask for a review, and get back a wall of text in the chat window. Sometimes useful. Often incomplete. Always requiring me to manually turn the output into something I could actually work with: a formatted document, a structured overview, a redline.

The first time I used Cowork for a contract review, I pointed Claude at an agreement, told it to assess against my playbook, and walked away. When I came back, there was a filled-out playbook with an updated overview of findings, categorised, sorted, and nearly ready to use. It can even draft a redline from there. Results on redlining are mixed but promising, and you'll still want human eyes before marking up anything you send out.

To give you a concrete sense: I fed it a complex agreement and my playbook. What came back were accurate assessments across the document, liability carve-outs flagged, audit rights gaps identified, DPA misalignments caught, termination assistance provisions questioned. Not perfect, but genuinely useful junior-level analysis on a document that would normally take hours to work through manually.

That moment shifted something in my thinking. Not because the AI was perfect (it wasn't) but because for the first time, the output didn't feel like raw material I had to shape. It felt like a draft from someone who understood the assignment.

The accuracy gap. The most significant difference I noticed compared to other AI tools came down to specific failure modes. In my experience, ChatGPT would hallucinate clauses that weren't there, miss sections entirely, or confidently reference provisions that didn't exist. With Cowork, I saw fewer hallucinated clauses, better section referencing, and more consistent references back to the actual document language. This is, of course, very important for legal work. An AI tool that's wrong 20% of the time isn't 80% useful: it's dangerous, because you spend your review time hunting for mistakes instead of doing analysis.

The chunking problem, solved. If you've tried reviewing a long agreement with any AI chat tool, you've hit the wall. The document is too large for the model to hold in its working memory all at once. In a regular chat, you're stuck: you can try pasting sections manually, but you lose the thread between them. Context degrades. The AI forgets what it read three sections ago.

Cowork handles this differently. It recognises when an agreement is too large to assess in one pass and segments it into manageable chunks automatically. It can even coordinate multiple agents working on different sections simultaneously, so the context window doesn't overflow and accuracy doesn't rot as the document gets longer. I watched it break a complex purchase agreement into parts, assess each one against the playbook, and stitch the findings back into a coherent overview. That's not something a chat window can do. One caveat: segmentation creates reconciliation risk. Cross-section dependencies (definitions, order-of-precedence clauses, limitation baskets) need to be re-checked in your final review.

The multi-skill coordination. This is where the "so what" becomes clear for daily legal work. Cowork doesn't just use one skill at a time. It reads a PDF using the PDF skill, builds an Excel summary using the Excel skill, and produces a Word document with redlines, all in the same task. It figures out which tools it needs and in what order. You describe the outcome you want; it decides how to get there.

For someone used to manually copying AI output from a chat window into a Word document, reformatting it, building a summary spreadsheet separately, and then cross-referencing everything, this felt like skipping three steps in a five-step process.

Why this is different from your ChatGPT use

I know what some of you are thinking: "I already use AI for contract review. This sounds incremental."

However, when you paste a contract into ChatGPT or a regular Claude chat, you're having a conversation. The AI reads what you give it, responds, and waits for your next instruction. Every task is one-shot. Every conversation starts from zero. If you want structured output, you have to ask for it explicitly. If the document is too long, you figure out how to split it.

Skills and (especially) Cowork change the operating model. Skills mean Claude already knows your preferences, your playbook, your formatting standards, without you repeating them. Cowork means Claude can plan and execute multi-step work autonomously, producing real files rather than chat text.

Regular AI chat is like texting a smart friend for advice. Cowork is like briefing a colleague and saying, "handle this and put the deliverables on my desk."

Both are useful. But they're useful for fundamentally different types of work.

What this means for your practice

If you're a lawyer who's sceptical of AI, I get it. Most of what you've seen probably hasn't been convincing, because most of it has been chat-based tools that produce mediocre output and require more cleanup than they save in time. Skills and Cowork don't eliminate the need for your judgment. But they do eliminate a significant amount of the manual labour between "AI gives me a response" and "I have something I can actually use."

If you're a lawyer who dabbles in AI but hasn't quite unlocked it, this might be the shift that makes it click. The difference between asking an AI to "review this contract" in a chat window and having it produce a formatted review document with an accompanying overview spreadsheet, that's a workflow change, not an incremental improvement.

And if you're already deep into AI and legal tech, look at what comes next. Anthropic's Legal plugin ships with seven skills: contract review, NDA triage, compliance workflows, meeting briefings, and more. But the real potential is in building your own. A legal ops team can take the contract review skill, encode their firm's specific playbook, add industry-specific clause types, and deploy it to the entire team. The same senior counsel judgment, applied consistently, every time.

I haven't built custom skills myself yet, that's a project for the coming weeks. But the idea that a firm's institutional knowledge could be packaged as a skill rather than trapped in a senior associate's head is worth paying attention to.

The honest limitations

I'd be doing you a disservice if I left out the rough edges.

Complexity has limits. When workflows involve many interconnected tasks, Cowork can struggle. It's excellent at structured, multi-step work, but "multi-step" has a ceiling. If you're expecting it to manage a full M&A due diligence process end to end, you'll be disappointed. It's a powerful tool for discrete tasks within a larger workflow, not a replacement for the workflow itself. So even though iteration is less important when prompting Cowork, there's still a place for chopping up your requests into blocks to keep things manageable.

Cost isn't trivial. Cowork requires a Claude Pro or Max subscription, with pricing that varies by plan and region. For a firm evaluating ROI, that's a real consideration, especially when industry data suggests that only 15% of AI decision-makers have reported measurable EBITDA lift so far, and Gartner projects that over 40% of agentic AI projects will be cancelled by end of 2027. This technology isn't free, and the returns aren't guaranteed.

Privilege and confidentiality remain unresolved. This is the big one for lawyers. The legal profession hasn't reached consensus on how AI tools interact with privilege, confidentiality obligations, and professional conduct rules. Depending on your practice area and jurisdiction, that may significantly limit where and how you can deploy these tools. Until clearer guidance emerges, caution is warranted.

It's not a replacement for legal judgment. Every skill in the Legal plugin explicitly states this. And it bears repeating, because the marketing around AI tools often implies otherwise. Cowork produces drafts. Good drafts, sometimes. But drafts that require a qualified lawyer to review, challenge, and approve. The moment you treat AI output as final work product is the moment you've created a liability. We're seeing that on both sides: clients who think they no longer need a lawyer's perspective, and lawyers who skip reviewing AI drafts properly. Both cost more time than they save.

Governance is coming whether you're ready or not. The EU AI Act and other legislative frameworks around the globe are coming into effect. Depending on the use case (particularly where AI influences decisions affecting individuals) some systems fall into higher-risk categories, with penalties reaching €35 million or 7% of global revenue. Firms should map their specific AI use cases against these obligations. And ironically, that governance expertise itself is something that can be encoded as a skill.

Where this is heading

The legal profession is approaching what the tech industry calls a build-versus-buy moment. When the base AI model performs at a high level, the value of many specialised legal tech products shifts. You're no longer paying for superior reasoning, you're paying for packaging, integration, and workflow structure. Skills allow legal teams to build that workflow layer themselves.

That doesn't mean every firm should start building tomorrow. But it does mean the technical barrier between "user" and "builder" has dropped significantly. Lawyers who understand how to orchestrate AI tools (scoping tasks, writing clear instructions, organising workflows, reviewing outputs critically) will have an advantage. Not because they've become technologists, but because they've learned to delegate effectively to a new kind of colleague.

52% of in-house legal teams are already using or evaluating AI for contract review. 64% expect to depend less on outside counsel because of AI capabilities they're building internally. Whether you find those numbers exciting or alarming probably depends on where you're sitting.

Either way, the direction is clear.

Nobody has this fully figured out yet. I certainly don't. But I've spent the past year testing these tools in real legal workflows (my own and those of the teams I work with) and the gap between "AI-curious" and "AI-competent" is smaller than most lawyers think. It mostly comes down to knowing where to start and what to skip.

If you're exploring this for your team and want to compare notes, my DMs are open. I find these conversations genuinely useful, they sharpen my own thinking as much as anything else.