AI prompting for lawyers: What works and what doesn’t

Many lawyers approaching AI start with the same assumption: that the system is in some sense “thinking.” That is the wrong mental model.

Large language models (LLMs) do not reason through legal problems the way lawyers do. They predict statistically likely text based on patterns in training data. Sometimes the results are excellent. Sometimes they are confidently wrong.

AI prompting for lawyers is the process of giving these systems enough structure, context, and direction to produce useful legal work product. The quality of the output depends less on the tool itself and more on the quality of the instruction.

AI behaves less like a search engine and more like a capable junior lawyer: fast, articulate, inconsistent, and highly sensitive to supervision. The lawyers getting real value from AI are not the ones using it most aggressively. They are the ones directing it most clearly.

AI isn’t magic, it’s predictive

LLMs are probabilistic systems. Ask the same question twice and you may get different answers. That unpredictability is part of what makes AI feel creative and human-like.

For lawyers, the implication is straightforward: AI output should never be treated as final legal work. Treat it as a first draft, a thinking aid, or a summarization tool. Review it the way you would review work from a junior lawyer.

Key takeaway: AI can be useful without being independently reliable.

The right mental model: AI as a junior lawyer

The junior lawyer analogy is the most useful way to think about AI.

A large language model:

• Has read widely
• Writes fluently
• Understands legal structure
• Responds quickly

But it also:

• Has no professional judgment
• Does not know when it is wrong
• Needs supervision

This is why vague prompts fail.

“Review this contract” is not a meaningful legal instruction. You would normally explain:

• Whose perspective matters
• What risks to focus on
• What output you want
• How detailed the review should be

AI requires the same structure.

Outputs also vary depending on framing. Small wording changes can produce different answers because the model works associatively. Keywords matter.

Key takeaway: Better legal instructions produce better AI output.

The hidden risk: AI will try to agree with you

Modern AI models are trained using human feedback. During training, humans preferred answers that sounded helpful and persuasive. The result is sycophancy: AI tends to validate the user’s framing.

That creates risk in legal work.

Ask: “Find me three cases supporting the argument that a limitation of liability clause was unenforceable due to gross negligence.”

The model may return plausible citations that do not exist. The danger is not only that the answer is wrong. It is that it sounds right.

A safer approach is adversarial prompting: “What are the strongest arguments against this clause being enforceable?”

Or: “Assume you act for the counterparty. Attack this position.”

That framing produces more balanced analysis and reduces hallucination risk.

Key takeaway: Do not ask AI to confirm your argument. Ask it to challenge it.

Why AI feels smart but has no memory

Large language models have enormous long-term knowledge and poor short-term memory.

Think Dory from Finding Nemo.

Large language models do not actually remember previous conversations. Each time you send a new message, the model rereads the entire current session from the beginning – including its own earlier responses. As conversations grow longer, earlier instructions can lose priority, output quality may degrade, and response times often slow down. In practice, starting a fresh session is often the better option.

Three practical rules follow:

• Keep unrelated matters in separate chats
• Start fresh sessions when output quality declines
• Be careful with persistent memory features

Some AI products now save “memories” about users. That can help with repetitive work, but it can also introduce unwanted assumptions across matters.

Key takeaway: Fresh sessions and clean context improve reliability.

Traditional document automation is often better for tasks requiring exact formatting, repeatable clause insertion, and precision.

Key takeaway: Use AI for drafting and structure. Be cautious where precision and authority matter most.

Why legal work is harder for AI than it looks

Most high-quality legal knowledge is not publicly available. It sits behind databases, inside firm precedents, or in practitioner judgment built over years.

That means AI is closer to a well-read generalist than a specialist lawyer. It understands legal drafting patterns and the structure of legal reasoning. It does not inherently know:

• Your client’s commercial priorities
• Unpublished market practice
• Your firm’s strategic approach

This is why legal AI products increasingly focus on governance, workflow, and document access rather than claiming to possess unique “legal intelligence.”

Input strategy: Less is often more

Lawyers often assume that more context produces better output. Often the opposite is true.

Large prompts and huge document uploads can overload the model. Important information buried in the middle may receive less attention than material at the beginning or end.

For contract review:

• Targeted sections usually outperform full-document uploads
• Summaries often work better than raw material
• Irrelevant boilerplate reduces quality

AI retrieval systems are also selective. They do not process every document equally. If the retrieval step misses an important clause, the model may never see it.

Key takeaway: Curated context beats document dumping.

The legal prompting framework

The Legal Prompting Framework is a six-step system for producing more reliable legal AI output.

Step 1: Be specific about the task

Weak promptReview this contract

Strong promptReview this SaaS master services agreement from the perspective of the customer. Identify clauses that limit the vendor’s liability in a way that would be unusual or unfavorable compared to market standard. List each clause, summarize the risk, and suggest alternative language.

Step 2: Provide relevant context

Weak promptDraft an employment termination letter

Strong promptDraft an employment termination letter for a senior employee in England and Wales being made redundant as part of a restructure. The employee has seven years of service. The tone should be formal but respectful. Include a short paragraph on the settlement agreement and do not include any admission of wrongdoing.

Useful context includes jurisdiction, deal type, risk tolerance, and client position.

Step 3: Define the output format

Weak promptSummarize the key risks in this term sheet

Strong promptSummarize the key risks in this term sheet as a table with three columns: Risk area, Description of the risk, and Recommended action. Limit each row to two sentences.

Step 4: Show the model what good looks like

Weak promptDraft an IP assignment clause

Strong promptDraft an IP assignment clause for a software development agreement. Use the following clause as a style reference: [paste example clause]. Cover moral rights waiver, assignment of future IP, and a warranty that the developer has authority to assign.

Examples improve consistency dramatically.

Step 5: Challenge, don’t validate

Weak promptIs this indemnity clause enforceable?

Strong promptI believe this indemnity clause is enforceable under English law. Identify the three strongest arguments a counterparty could make to challenge enforceability and assess how likely each is to succeed.

Step 6: Build and reuse a prompt library

Teams getting consistent value from AI rarely start from scratch. They maintain reusable prompts for:

• Contract review
• Due diligence
• Board summaries
• Client updates
• Clause drafting

A useful structure is:

1. Task
2. Context
3. Output format
4. Example
5. Counter-analysis instruction

Key takeaway: Good prompting is structured legal supervision translated into instructions.

Conclusion: The competitive advantage is clearer thinking

AI does not replace legal expertise. It amplifies it.

A lawyer who gives vague instructions gets vague output. A lawyer who defines the task clearly, provides context, and critically evaluates the result gets genuinely useful work product.

The competitive advantage will not belong to whoever adopts AI first. It will belong to the lawyers who become best at directing it.

That is already a core legal skill: asking the right question, in the right way, with the right context.