Legal AI prompts for lawyers: a practical prompt library

A prompt structure that works

1. Role: 'You are assisting a commercial litigator in [jurisdiction].'
2. Task: one clear action, such as 'summarise', 'list issues' or 'redraft'.
3. Materials: paste the text, and say it is the only source to use.
4. Format: bullets, a table, a 150-word email, headings.
5. Limits: 'If the text does not say, write NOT STATED. Do not cite cases.'

The last line matters most. Telling the tool not to cite authority and to flag gaps cuts down the risk of fabricated content, though it never removes the need to check.

Prompts to adapt

Summarising

• 'Summarise the text below in 10 bullets for a partner who has not read it. Then list any dates or deadlines mentioned. Use only this text.'
• 'List the obligations of [Party A] in the clause below, one per line, and flag ambiguous wording.'

Drafting

• 'Redraft this paragraph in plain English at the same legal effect, then list any change that might alter meaning.'
• 'Write a polite 120-word email to a client explaining that we need the documents listed below by [date]. Tone: warm, direct.'

Issue spotting and preparation

• 'From the facts below, list questions I should ask the client in a first meeting. Do not give legal conclusions.'
• 'Suggest counter-arguments to the position below and say what evidence would support each.'

Keeping the library in Notion

Create a simple database with columns for prompt name, use case, practice area, the prompt text, last tested date and 'approved for client data' (yes or no). Review it quarterly and retire prompts that gave poor results. A shared library helps with supervision because you can see what the team is actually asking tools to do.

The Custom AI Agents page in LawyerOS Pro with AI ($39, one time) is a place to keep this kind of instruction set for Notion's own AI features. LawyerOS itself is a template and contains no AI model. See /post/custom-ai-agents-for-lawyers.

Safety checklist before sending any prompt

• Have I removed names, addresses and file numbers, or confirmed the tool's terms permit them?
• Am I asking for analysis I will check, not authority I will rely on?
• Is the output going to a client or court? If so, who reviews it?
• Does my regulator or the court have guidance on AI use here?

Frequently asked questions

What is the best prompt format for lawyers?

Role, task, materials, format and limits. Always tell the tool to use only the text provided.

Can I put client names in a prompt?

Only if the provider's terms and your confidentiality duties allow it. Otherwise anonymise.

Should I ask AI for case citations?

Avoid it. Models can produce citations that look right but do not exist. Find authorities in a proper legal research source and ask AI to work from those.

Do prompts give the same answer every time?

No. Outputs vary, which is another reason to review everything.