Free resource · no email, no sign-up

Claude for Lawyers: every prompt from the video

Every lawyer I talk with tells me the same thing. Claude makes bad decisions, it feels clunky, and it doesn't get the job done fully. These are the 5 strategies from the video that fix that, in the order they appear, with every prompt ready to copy into Claude.

Alejo Pijuan · Chief AI Officer, Realize Partners

1
Strategy 1

You decide, not Claude

Claude's problem is that it makes the decisions the human should be making. Add one snippet to the end of any request and Claude asks you what it needs, then stops at every judgment call and shows you the options instead of picking one.

The second one brought in your judgment. It brought you questions and can even bring you questions while it's doing the research. All you have to do is ask.

In the video · strategy 1

The snippetStart here today
Don't decide for me. Ask me strategic questions that give you enough context about the case to get the job done right. When you hit a judgment call, stop, show me the options with the exact source text, tell me what you couldn't verify, and ask me.

The example from the video

The same personal injury case, twice. Run B only adds the snippet. Run B asked 4 questions first and wrote the memo around the answers.

Run A · the task alone
Our client was rear-ended at a red light in Florida. The defense says our driver stopped short. Write me a memo on what a defendant has to produce to rebut Florida's rear-end presumption of negligence, and how comparative fault affects it.
Run B · plus the snippet
Our client was rear-ended at a red light in Florida. The defense says our driver stopped short. Write me a memo on what a defendant has to produce to rebut Florida's rear-end presumption of negligence, and how comparative fault affects it.

Don't decide for me. Ask me strategic questions that give you enough context about the case to get the job done right. When you hit a judgment call, stop, show me the options with the exact source text, tell me what you couldn't verify, and ask me.
Still check every case it cites. Open each one before you rely on it. Claude does the finding, you do the deciding.
Your first step today

Paste the snippet at the end of the next thing you ask Claude for.

2
Strategy 2

Context beats prompting

You've heard for years that it's all about prompting. Something matters more: what Claude can read before it answers. One of the big reasons Claude feels dumb is that it forgets. Put your documents where it can reach them, then tell it to go look.

No matter how good your prompt is, if the context is not something Claude has access to, you're going to get a bad result.

In the video · strategy 2

The promptStart here today
Before you answer, look through the documents and emails you have access to for anything relevant to this, and tell me what you read.

The example from the video

  1. 1Give Claude your folder first: in Cowork, click Add folder.
  2. 2Run the same request twice. Run B only adds the prompt above.
Run A · the task alone
Review the vendor agreement in our folder.
Run B · plus the prompt
Review the vendor agreement in our folder. Before you answer, look through the documents and emails you have access to for anything relevant to this, and tell me what you read.
Run A got

A breach, some recommendations, nothing of substance.

Run B read 4 documents first

The firm's rules, the agreement itself, the last 2 weeks of partner notes, and the firm's profile.

Your first step today

Give Claude access to one folder of real documents, then add the prompt to your next request.

3
Strategy 3

Teach Claude a skill

If you review one vendor agreement, you'll probably review many. A skill is simply a repeatable workflow: you teach it once, and every time you type a slash and its name, Claude knows exactly what comes next. It asks you for the nuance, and it gets better every time you correct it.

Turn work you just did into a skill

Say this at the end of any conversation that went well.

From work you just did
Create a skill based on the work that we just did. Identify the pros and cons and create a repeatable skill.
Or from scratch
Create a skill that can help me review vendor agreements.
Then use it in a new conversation
/contract-review
Review the vendor agreement in the folder.
You don't need to add the context prompt or the decision snippet again. They already live inside the skill.

Your legal OS: the one skill you run every session

These skills are very powerful, but at the end of the day, they get a single narrow job done.

In the video · strategy 3

A legal operating system is simply the company's brain. It knows your firm and brings you in for the decisions that are yours. Run it every time you sit down to work.

Build your legal OSStart here today
Create a skill for my own legal operating system. It should know everything about our firm. It should ask me how I want it to interact with me, and it should bring me in to make decisions as the human at the right point in the task. Ask me any strategic questions to get the job done right.
Run it at the start of every session
/legal-os
When it asks you
Start strict: have it bring you in more often than not, and loosen it later. Ask it to handle your firm's memory too.
Build this one on Fable. In the video I switched from Opus to Fable for it, because every task after it builds on it. For daily work, Opus is still a really good model.
Your first step today

Paste the legal OS prompt into a new conversation and answer its questions.

5
Strategy 5

Use Claude safely

Don't end up like the lawyers getting fined over AI-hallucinated cases. There's a lot of nuance to staying compliant and safe, so it comes down to 3 decisions. If you don't want any headaches, the short answer is Claude Enterprise, set up properly.
The one thing never to do: paste full client emails or unredacted documents into a personal Claude account.

The 3 decisions

If you have
Zero tolerance for a privilege or privacy problem
Claude Enterprise, or a private model that redacts first

Either nothing sensitive leaves the building, or the firm controls how long Anthropic keeps it.1

If you have
More than 20 seats, lawyers or staff
Claude Enterprise

No training on your data, admin controls, and you set the retention period yourself.2

If you have
Fewer than 20 seats
A private model redacts on your computer, then Claude

The data is clean before Claude sees it. Enterprise becomes optional.3

The table

The small numbers point to the sources underneath.

How you use ClaudeTrains on youKept how longFirm controlsClient work
Personal account
Free, Pro, Max
Unless you turn it off4Up to 5 years5NoneNo
Claude Team
No6Not published7Admin, SSOThe floor
Claude Enterprise
No6You set it2FullYes, once set up
API with zero retention
a build, not the chat app
No6Nothing stored8YoursYes
Private model redacts, then Claude
runs on your own computer
No9Nothing sensitive leavesYoursYes

If you go the private model route

Run this on a model that lives on your own computer, read the result yourself, then give the redacted version to Claude.

Redaction prompt for a local model
Redact every name, date of birth, address and claim number in this document. Replace each one with a label like [NAME 1], [DATE OF BIRTH], [ADDRESS], [CLAIM NUMBER]. Return the full document with nothing else changed.
Your first step today

Ask whoever bought Claude at your firm.

The questionStart here today
Which Claude plan is our firm on, and is training on our data turned off?
The full workshop
Is Claude for Legal safe? It depends on your plan and setup →
Sources for the table (9)
  1. A private model on your own machine never sends the original anywhere. On Enterprise the firm's admin sets the retention period, and if nobody sets it the data is kept indefinitely, so "set up" is not optional. Anthropic Help Center, custom data retention for Enterprise.
  2. Enterprise retention is a setting you choose, not a negotiation: Organization settings, Data and Privacy, minimum 30 days. Projects default to indefinite, and a project's retention beats the chat retention for chats inside it. Same page.
  3. Redaction is the control, so a person reads the result before it goes. Our own setup is Ollama with a local model plus Hermes on the Mac, which is not an Anthropic product.
  4. Personal accounts: turn off "Help improve our AI models" in Settings, Privacy. How to change the setting. The Consumer Terms and the Privacy Center read slightly differently on the default: Is my data used for model training (consumer).
  5. Up to 5 years, de-identified, in the training pipelines if training is allowed; no standard consumer retention period is published today. Delete a chat and it leaves the back end within 30 days. Pressing thumbs up or thumbs down on any plan keeps that whole conversation for up to 5 years. How long do you store personal data.
  6. Commercial plans: "By default, we will not use your inputs or outputs from our commercial products" to train. Is my data used for model training (commercial). On Team and Enterprise an owner can switch off the "Rate chats" button.
  7. Team: no published retention length and no custom retention setting. The Primary Owner can export the workspace, so Team is safer from the vendor and not private from your own firm. How long do you store my organization's data.
  8. Zero data retention is negotiated with Anthropic, approved per organization, and covers the API and Claude Code, not the claude.ai Enterprise chat app. Safety classifier results are still kept, and the newest covered models carry a mandatory 30-day safety retention. What a ZDR agreement applies to, Covered Models retention. The plain API without zero retention deletes inputs and outputs within 30 days.
  9. The local model never leaves the building, so there is nothing to train on. After redaction, the row you picked above governs what Claude receives.

Every layer you add is another set of terms: Claude inside somebody else's platform means their privacy policy and Anthropic's. We are not lawyers and nothing here is legal or ethics advice; your state bar's guidance governs. Anthropic's pages checked 21 September 2026.

Hit a wall?

30 minutes, pro bono, and we get you unstuck

We sit down, look at where it went wrong and what happened, and solve it together. It's not a sales call. Or send me a screenshot of the issue on LinkedIn and I'll write back.

Book 30 minutesMessage me on LinkedIn