Every attorney who has set up Claude for a firm has asked some version of the same question: is this the right way to do it. Wrong question. There is no single right way to set up Claude — not for a solo practice, not for a firm with associates, not for a coaching client three weeks into their first suite. There are two questions worth asking about every piece of work before touching a setting, and the “right” setup is just an honest answer to both of them, every time.

Question one: what’s going in. What can Claude actually see for this piece of work, and where does it come from. Typing into a chat puts in exactly what you typed, for that conversation, and nothing more unless memory picks it up. Uploading a file to a chat is a one-time document informing that one conversation. A project’s knowledge base is different — persistent, available to every chat inside it, and to everyone the project is shared with. A connector to an inbox or a practice management system reads live data without anyone uploading anything, and depending on how it was set up, it may still be reading long after the person who granted it forgot it existed. Cowork mounting a folder on a hard drive is the deepest version of input there is — the actual matter file, not a copy of it. None of these are wrong. They’re just different amounts of exposure, and the mistake is not choosing the amount on purpose.

Question two: what’s coming out. What does Claude produce, who else is allowed to see it, and how do you get your own material back or make it stop existing anywhere. A reply in a chat window is the lightest output there is — text that exists only in that conversation until someone copies it somewhere. A file Cowork writes to a mounted folder is the heaviest — it lands on the drive the moment the task finishes, not as a draft waiting for review. Sharing a project decides who besides the person who built it can see everything inside it, from private to the whole firm. Exporting data — the one path Claude offers, account-wide, a 24-hour download link, nothing per-conversation — is how you get your own information out of the system entirely rather than just out of view. Deleting a conversation removes it from view immediately and finishes clearing it from the back end within 30 days. None of these should feel like an afterthought either, because the input question and the output question are connected: what you put in becomes exactly what everyone named by the output setting can see.

That connection is where firms actually get hurt, not the individual settings. A shared reference project — the caselaw library, the drafting templates, the standing instructions every attorney should be pulling from — is built to answer the output question generously: share it broadly, because nothing in it should be a secret from the people using it. That’s correct, until someone answers the input question carelessly and uploads one client’s financial disclosures into that same project. Now everyone with view access to the firm’s training library has a specific client’s sensitive documents, and no individual setting was wrong. The two questions were just answered in the wrong order.

A live matter answers both questions the opposite way, and that’s also correct — for that work. A Cowork project pointed at the actual matter folder wants deep input: the whole file, not a curated summary, mounted directly rather than uploaded piece by piece. What it should not want is broad output. It stays on the machine of whoever’s working the file, its memory scoped to that one case, never shared the way the firm-wide project is. When the matter closes, archiving that Cowork project actually deletes its setup — name, instructions, links, memory — which is not how archiving works on a shared project, where it’s fully reversible. Same word, different product, different outcome, and worth knowing before a matter closes rather than after.

And some work should answer both questions with nothing at all. A quick question involving something genuinely sensitive doesn’t need a project or a connector — it needs incognito, decided before the first word is typed. Nothing saved to history, nothing added to memory, nothing surfaced later when Claude searches old chats for context. It’s the one control on this list that has to be chosen in advance. Everything else can be cleaned up after the fact to some degree; this one can’t.

None of this is a reason to avoid building any of it. It’s a reason to ask both questions on purpose, for the specific piece of work in front of you, instead of copying a setup that was right for something else.

We put together a full breakdown of every control on both sides — nine ways data goes in, eight ways it comes out, and the four mistakes that show up when the two questions get answered by accident — as an interactive reference guide: Data In, Data Out.

Talk to us:

1. Book a free exploration call — 15 minutes, no pitch, just whether this fits your practice.
2. See a free private demo — 30–45 minutes, built around your actual workflow.
3. Browse the systems and see what an AI paralegal actually does, system by system.

AI built for law firms. Powered by real lawyers.