Tov Law Grow Your Autonomous Law Firm Free AI Resources for Lawyers
§ 42 · AI assistants

Claude Projects for Lawyers: Your Firm's Knowledge, Reused

A Claude Project is a workspace that remembers your firm. You load your templates, your best past work, and your instructions once, and every conversation started inside it begins with that context already in place. For legal work, which is repetitive by nature, this is the difference between a tool you demo and a tool you use.

Most lawyers who try AI and conclude it writes like a first-year associate are experiencing a context problem, not a capability problem. They opened a blank chat and asked for a document, so they got the average of everything ever written. Load your own work and the output starts sounding like your firm, because it is now imitating your firm.

What goes into a Project

Three things, and the order matters.

Custom instructions come first. This is a paragraph or two describing who you are, what you practice, your jurisdiction, your tone, and any hard rules. Mine include things like never inventing a citation, always asking for missing facts rather than assuming them, and writing in plain language rather than legalese.

Documents come second. Your template for the document type, and three to ten examples of that document done well. Resist the urge to upload everything. Ten excellent demand letters produce better output than a hundred that include your worst ones, because you are teaching by example and every example counts as a vote.

Then the conversations themselves, which is where case facts go.

That last point is the confidentiality structure worth adopting: the Project holds templates and redacted exemplars, and live client facts stay in individual conversations. It keeps the persistent layer clean.

The six Projects worth building first

Demand letters, if you are on the plaintiff side. Template plus your best examples plus your tone. This is the highest-return one and it pairs with the workflow in AI demand letter software.

Client updates. Load the update format you want and a few good examples. Status emails are the thing firms are worst at and clients complain about most, and they are almost pure formatting.

Intake summaries. Load your intake criteria and what a good summary looks like. Paste in the notes from a call, get back a structured summary against your own qualifying questions.

Discovery responses. Load your objections library and past responses. This one saves the most hours per week in a litigation practice.

Firm SOPs. Load your written procedures so the Project can answer staff questions about how your firm does things. This one quietly reduces interruptions to your managers more than anything else on the list.

Marketing. Load your best posts and your brand voice, so content stops sounding like it came from a template mill.

Writing the instructions

The instructions are where most Projects fail, usually because they are too polite and too vague.

Be specific about what you never want. “Never invent a case citation. If you are unsure whether a case exists, say so.” Be specific about missing information. “If a fact needed for this document is missing, ask me for it rather than writing around it.” Be specific about voice. “Write at an eighth grade reading level. Short sentences. No legalese where a plain word exists.”

State the jurisdiction and practice area explicitly, every time. A model told it is drafting for a California lemon law practice behaves differently than one guessing from context.

Where Projects fall short

They do not connect to your systems. A Project cannot read your case management data or your calendar. It knows what you uploaded and what you paste. If you want the AI reaching into Clio or your billing system, that is a different mechanism, and MCP with the Clio integration is the path.

They go stale. A Project loaded with 2024 templates will keep producing 2024 templates. Someone has to own refreshing them, and in practice nobody does unless it is assigned.

And they drift across a firm if everyone builds their own. Six attorneys with six personal demand letter Projects is six different house styles. Build them once, share them, and put one person in charge.

Do this today

Pick the document your firm writes most often. Create a Project. Upload your template and your three best examples of that document. Write four sentences of instructions covering practice area, jurisdiction, tone, and the rule that it must ask for missing facts instead of inventing them.

Then run a real matter through it and compare against what you would have written. Setup takes about 20 minutes. If the output is not close, the fix is almost always better examples rather than better prompting.

Questions lawyers ask

What is a Claude Project and how is it different from a normal chat?
A Project is a workspace with persistent knowledge attached. You upload documents and write custom instructions once, and every conversation inside that Project starts with all of it already loaded. A normal chat forgets everything when you close it. For repeated work like demand letters or client updates, that difference is the entire value.
How many documents can you put in a Claude Project?
Enough for a practice area's worth of templates and exemplars. The practical limit is usefulness rather than capacity: loading 200 documents makes output worse, not better, because the signal gets diluted. Three to ten excellent examples of the specific document you want beats a hundred mediocre ones.
Is client data safe in a Claude Project?
On business and enterprise plans, conversations are not used for training and retention is configurable. On consumer plans, read the current terms yourself before uploading anything client-identifying. A safe default many firms adopt is to load templates and redacted exemplars into the Project, and paste live case facts into individual conversations.
Do Projects work better than a long custom prompt?
Yes, for anything you do more than twice. A custom prompt has to be pasted every time and drifts as people edit their copies. A Project holds the instructions centrally, so the whole firm gets the same behavior and you fix it in one place.

Go deeper