Practical Applications and Patterns

Mitchell Adams

20 Practical Applications and Patterns

Effective use of AI in legal practice requires more than understanding how to construct prompts — it demands an appreciation for how these tools can be integrated meaningfully into legal workflows. While generic prompting techniques provide a foundation, legal work presents unique challenges and requirements requiring specialised approaches.

This section connects theoretical knowledge with practical application, offering examples and prompt patterns to assist with legal tasks. Through carefully selected use cases and example prompts, it explores how to leverage large language models for various tasks. These patterns serve not just as templates to be followed but as frameworks that can be customised to meet the needs of legal practice.

Understanding Legal Tasks and Their AI Suitability

Before integrating generative AI into legal practices, it is essential to assess which tasks are best suited for AI assistance. This involves distinguishing between routine and complex tasks and evaluating the readiness of legal data for processing by generative AI.

Legal tasks can be broadly classified into routine and complex categories. Routine tasks are repetitive and require minimal discretion, such as data entry, document categorisation, and straightforward reviews. These tasks are ideal candidates for AI automation because they often follow predictable patterns that a model can learn and execute.

On the other hand, complex tasks involve higher levels of judgment, expertise, and tailored solutions, such as crafting legal arguments, negotiating deals, or advising clients on strategy. While generative AI can support these tasks by providing information and preliminary analysis, it generally requires a deeper human touch to interpret nuances and make final decisions.

Additionally, the success of models in legal tasks depends significantly on the availability and quality of data. AI models require a substantial amount of data to learn from; therefore, understanding what data models are trained on can help determine which models to use (or even augment RAG, as described above). Furthermore, the ethical and privacy implications of using sensitive legal data must be considered. This includes ensuring that data is anonymised or removed where appropriate and that AI use aligns with regulatory compliance requirements related to data security and client confidentiality.

The remaining sections of this chapter set out examples and prompt patterns under the broad categories of:

Text transformation
Legal research
Communication
Document assembly
Legal writing
Training
Creativity

Text Transformation

Legal work often involves processing vast amounts of textual information — from contracts and case law to statutes and regulations. AI-powered text transformation enables lawyers to extract, summarise, compare, and restructure legal text efficiently. Through well-crafted prompts, practitioners can convert dense legal documents into actionable insights, identify key provisions across multiple agreements, or transform unstructured legal opinions into structured data. This capability saves time and enhances comprehension of complex legal materials, allowing legal professionals to focus on analysis rather than information processing. The following section provides prompt templates tailored to various text transformation tasks.

Summarisation

Large language models such as ChatGPT excel at summarising information. However, the model’s performance largely depends on the prompt you provide. As discussed previously, providing a prompt such as “Summarise the Trade Marks Act 1995 (Cth) for me” relies on the model being trained on the piece of legislation or forms part of the corpus of legal information in an RAG system.

Have a look at the results of this prompt with ChatGPT 4o here: https://chatgpt.com/share/67be73aa-0664-800d-84d9-821a32064627

This is a good example of how models like ChatGPT are predominantly trained on American data. The output spells out trad emark in the American way (as one word) and hallucinates by introducing concepts like ‘honest concurrent use’ in trade mark infringement.

By employing the techniques mentioned above, a more effective prompt will guide the model’s output. To maximise the benefits of having a model summarise lengthy texts, provide the necessary context (including examples), offer a template that outlines your preferred format, and include placeholders for the types of information you wish to see.

One approach offered by CaseMark is the prompt below.^[1] No, that is not code you can see in the prompt. The syntax in the prompt is Markdown, a markup language that specifies the structure and formatting of a document. Markdown is a good tool for prompting a model to produce responses in a desired format (See Fundamentals to Prompting for more information).

As a legal professional, analyse the securities act of 1933 and generate a detailed summary with 3-4 highlights for each of the most important sections with important keywords, people, numbers, and facts in this format:

# {title here}

### {section title here}

{START_DETAILS_SECTION}

{summary of the section with important keywords, people, numbers, and facts}

– {first point}: {short explanation with important keywords, people, and facts}

– {second point}: {same as above}

– {third point}: {same as above}

<!– a fourth point if warranted, and so on –>

{END_DETAILS_SECTION}

### {second section here}

{START_DETAILS_SECTION}

{summary of the section with important keywords, people, numbers, and facts}

– {first point}: {short explanation with important keywords, people, and facts}

– {second point}: {same as above}

– {third point}: {same as above}

<!– a fourth point if warranted, and so on –>

{END_DETAILS_SECTION}

### {third section here}

<!– and so on, as many sections and details/summary subpoints as warranted –>

End with other questions that the user might want answered based on this source:

### Further information

– {question + answer 1}

– {question + answer 2}

– {question + answer 3}

All words in brackets must replaced by the summary of the content.

Only draw from the source content, do not hallucinate.

Only output the response in the prescribed format with NO additional commentary.

Answer the questions included in the “Further information” section and display the answer alongside the question.

This prompt provides a solid template for the model to complete, outlining the desired sections with headings and instructions. However, it assumes that the model is familiar with the Securities Act of 1933 and is likely to produce results relevant to American jurisdiction. A more effective prompt would include the text of the most current, in-force legislation (to prevent issues with knowledge cutoffs) and impose restrictions on the jurisdiction

David Colaursso presents the following prompt pattern, emphasising summarisation and simplification.^[2] He uses variable placeholders in his prompts. Notably, double curly brackets denote predefined variables; in this case, they indicate the location where the text to be summarised should be copied into.

Provide three short summaries of the above text in increasing complexity. The first should rephrase the text such that a 12-year-old could understand it. The second should be for a general audience, written in plain language without jargon and taking time to explain potentially unfamiliar concepts. The third should match the level and tone of the source material, including nuance and technical detail.

If asked any follow-up questions, use the above text, and ONLY the above text, to answer them. If you can’t find an answer in the above text, politely decline to answer, explaining that you can’t find the information. You can, however, finish a thought you started above if asked to continue, but don’t write anything that isn’t supported by the above text. And keep all of your replies short!

Now give your three increasingly complex summaries, starting with a super simple version. That is, they should be aimed at a: (1) 12-year-old, (2) general audience, and (3) wonk.

This prompt is not only beneficial for summarising text, but it also includes a solid constraint to prevent hallucination. Specifically, the model is instructed to consult only the given text and not to go beyond it. In any other case, the model should produce a commentary or note that it could not find the answer or decline to answer. Try using the prompt and see if it goes beyond any supplied text.

Question and Answer

Generative AI tools can also be prompted for document question-and-answer (Q&A) sessions. This task would involve uploading the document and querying the specific document. Such a task has performed well in benchmark testing.^[3] Vals noted that: ‘having an accurate, fast question-answerer can save lawyers a significant amount of time and effort, as they no longer need to manually search through documents to find the answers to their questions. For long, complicated legal documents, this can be especially valuable.’

To achieve good results, augment RAG by uploading the document to a chosen tool and asking it a natural-language question. For example, “What do I do if an employee says they can’t work on certain weekdays for religious reasons?”^[4] Alternatively, an off-the-shelf solution can be purchased for Q&A. For example, the Josef Q product or Checkbox helps you build a Q&A tool based on a corpus of uploaded documents.

Classification

Building on text summarisation, LLMs excel at classifying or categorising text provided to them. Sentiment analysis of text was identified early on as a strong use case for generative AI systems. An example prompt for sentiment analysis might be:

# Sentiment Analysis Task

Analyse the sentiment of the given sentence and classify it as positive, negative, or neutral.

## Instructions:

First, identify the key emotional words or phrases in the sentence
Consider the context and implied meaning
Determine the overall sentiment (positive, negative, or neutral)
Explain your reasoning with specific references to the text

## Examples:

Example 1:

Sentence: “Same S##t, different day”

Analysis:

Key phrases: “Same S##t” (negative expression of frustration), “different day” (implies repetition)
Context: Expresses frustration with recurring situations or tasks
Sentiment: Negative
Reasoning: The phrase indicates boredom and frustration with repetitive circumstances. The censored profanity reinforces the negative emotional tone.

Example 2:

Sentence: “This product has changed my life. I am so much more productive now”

Analysis:

Key phrases: “changed my life” (positive transformation), “much more productive” (positive improvement)
Context: Describing the beneficial impact of a product
Sentiment: Positive
Reasoning: Both statements indicate significant improvement in the speaker’s life. The phrases “changed my life” and “more productive” express clear satisfaction and appreciation.

## Your Task:

Sentence: [INPUT SENTENCE]

Analysis:

Key phrases:
Context:
Sentiment:
Reasoning:

This prompt is an example of manual chain-of-thought prompting, as it provides a framework for analysis.

Asking a model to classify text can occur in a couple of ways: you can either provide the model with the desired categories or ask it to generate its own. The model will perform better for this kind of use case when you include some examples as part of the context. Therefore, if you have pre-defined categories, supply the model with an example text and its categorisation.

A similar prompt could be:

# Legal Document Classification Task

Analyse the given legal document excerpt and classify it according to document type, purpose, and relevant legal area.

## Instructions:

First, identify key structural elements and language in the document
Note the formal characteristics (formatting, terminology, clause structure)
Determine the document type and primary purpose
Specify the relevant legal practice area(s)

## Examples:

Example 1:

Document Excerpt: “WHEREAS, Party A (‘Licensor’), a corporation organized under the laws of Delaware, owns certain intellectual property; and WHEREAS, Party B (‘Licensee’) wishes to obtain rights to use said intellectual property in connection with its business operations; NOW, THEREFORE, in consideration of the mutual covenants contained herein, the parties agree as follows:”

Analysis:

Key elements: “WHEREAS” recitals, “NOW, THEREFORE” transition, parties designated as “Licensor” and “Licensee”, reference to intellectual property
Formal characteristics: Formal recital structure, consideration clause, party designations
Document type: Licensing Agreement
Legal practice area: Intellectual Property Law

Example 2:

Document Excerpt: “Notice is hereby given that a meeting of creditors of XYZ Corporation (the ‘Debtor’) will be held on June 15, 2025 at 10:00 AM at the offices of Smith & Jones, LLP. The purpose of this meeting is to consider the Debtor’s proposed plan of reorganisation dated May 1, 2025. Creditors wishing to vote on the plan must submit completed ballots by June 10, 2025.”

Analysis:

Key elements: Notice to creditors, meeting details, reference to a debtor, plan of reorganisation, voting procedures
Formal characteristics: Notice format, specific dates and procedural instructions, formal designation of parties
Document type: Creditors’ Meeting Notice (Bankruptcy Proceeding)
Legal practice area: Bankruptcy Law

## Your Task:

Document Excerpt: [INPUT LEGAL DOCUMENT EXCERPT]

Analysis:

Key elements:
Formal characteristics:
Document type:
Legal practice area:

A team of researchers explore the performance of LLMs using different prompt engineering approaches in the context of legal text classification. Parizi et al. investigated the performance of the OPT family of models (with equivalent capabilities to GPT-3).^[5] This research tested various prompting strategies on legal texts and found that standard approaches often fail in legal contexts due to specialised terminology and document complexity. Few-shot prompting consistently outperformed basic prompting, while chain-of-thought reasoning unexpectedly hindered performance in legal scenarios. The most effective technique is to combine providing examples with ‘activation fine-tuning’.

When classifying legal documents, focus on crafting clear, completion-style prompts (e.g., Classify the given document with only one label from the following list: [Intellectual Property, Civil Rights, Criminal Procedure, etc.]) with relevant examples rather than open-ended questions.

Extraction

Text transformation using AI offers powerful capabilities for extracting structured, actionable information from unstructured legal documents. Capabilities include specific information retrieval (e.g. key terms from contacts, legislative mentions in cases) and named entity recognition (extracting people, organisations, locations and dates). LLMs also can complete context-aware extraction of contractual elements (obligations, rights, conditions), detection of relationships between entities,^[6] temporal sequence reconstruction (i.e., chronologies), and conversion of unstructured conversation into organised factual timelines.^[7] Once extracted, structured data can then be analysed, compared and operationalised.

For example, Damien Riehl created this prompt to extract SALI (Standards Advancement for the Legal Industry) tags from text.^[8]

From the text above, please extract verbatim text snippets. Place each snippet into a Markdown table. Extract only verbatim snippets. If you’re not sure, do not answer.

Below is the syntax for creating that Markdown table:

< COLUMN>: <SNIPPET EXAMPLES>

Within any column, if a cell has multiple snippets, then make it a new row. Using this syntax, below are the instructions for creating that Markdown Table:

AREAS OF LAW: Patent Law, Corporate Law, Finance and Lending Law

INDUSTRY: Health Care, Information, Insurance

LEGAL CONCEPTS: Appeals, Advice, Debt Financing, Litigation, Breach of Contract, Negligence, Patent Infringement, Trade Secret Infringement, Copyright Infringement

ACTORS: Creditor, Debtor, Licensor, Licensee, Plaintiff, Defendant, Guardian, Inventor, Author

ASSETS: Real Estate, Office, School, Goods, Stocks, Bonds, Equities, Cash, Trademarks, Patents, Copyrights, Trade Secrets

FORUM: SDNY, N.D. Cal., Eastern District of Texas, Florida Supreme Court, Illinois Court of Appeals

LEGAL ENTITY: Private Company, Public Company, S Corp., C Corporation

AUTHORITIES: Cases, Constitutions, Legislation, Statutes, Regulations, Treatises, GDPR, DMCA

LOCATIONS: New York, California, Texas, United States, China, Russia, EU, Germany, UK

GOVERNMENTAL BODY: Federal Trade Commission, New York Division of Human Rights, Department of Justice, Department of Labor

Generative AI tools are anticipated to excel at extraction tasks when given short texts or prompted to extract a specific section or clause. The Vals Legal AI benchmark report recently assessed the tools CoCounsel, Vincent AI, Harvey, and Oliver on data extraction tasks.^[9] Each performed well in this task. However, the authors indicated that there are additional considerations when tasking a Generative AI tool to extract information from larger documents. In the context of extracting information from contracts, the authors noted:^[10]

(1) Applications make mistakes when there are multiple clauses that need to be extracted for a single question. The tools often stop at one relevant clause without producing an exhaustive answer. For the question, “Extract any text related to disclaimers of warranties,” the expected answer had three distinct clauses that should have been extracted. All applications found at least one of the clauses but no tool was successful in finding all three. A better tool may find more relevant clauses, but in their current state, the output of generative AI tools should be viewed as a starting point for further work by the lawyer.

(2) Some tools did not understand terms of art to the degree required to produce an accurate answer (see the example below). Some concepts, such as Most Favored Nation or the total available credit “basket,” are used in extraction prompts for documents, but are descriptive of types of clauses, rather than being present verbatim in the text itself. The better tools were able to interpret the semantic meaning of these phrases in their effort to find a text fragment.

(3) The hardest questions for the AI tools were those in which the model had to “read” more than one document and produce extracted text across all documents. One such question was, “Extract from each of the agreements the Security Deposit at the start of the Lease term, excluding any Letter of Credit Amount. Put the amounts in a table where each lease is on its own row.”

One of the key use cases for automated extraction is converting unstructured judicial decisions into structured data. This use case may apply to the empirical analysis of judges’ decisions, which could reveal patterns in successful and unsuccessful outcomes.

In 2024, researchers from Cambridge used GPT-4 for information extraction from UK Employment Tribunal (UKET) cases.^[11] The team aimed to extract the following data from each case:

claims made by the parties,
references to legal statutes and precedents,
general case outcomes summarised as one of four labels (namely, “claimant wins”, “claimant partially wins”, “claimant loses”, and “other”),
detailed orders and remedies, and
reasons for the decision.

Additionally, the team sought to analyse three of these extracted features to facilitate the development of a tool capable of predicting the outcomes of employment law disputes.

The prompting techniques they used included:^[12]

You are a legal assistant. Your task is to read through the court decisions that I will send you, and extract the following information for each input:

Facts of the case of the specific court decision;
Claims made in the specific court decision and considered in the specific court decision. Do not include any claim which has already been decided in any previous decision;
Any references to legal statutes, acts, regulations, provisions and rules, including the specific number(s), section(s) and article(s) of each of them, and including procedural tribunal rules;
References to precedents and other court decisions;
General case outcome;
General case outcome summarised using one of the following four labels – “claimant wins”, “claimant loses”, “claimant partly wins” and “other”. Note that the label “other” is to be reserved for situations in which the result cannot be determined or where the outcome cannot be described in terms of winning or losing (e.g., an evidence collection);
Detailed order and remedies;
Essential reasons for the decision (procedural and substantive). If there are multiple claimants or respondents, extract the case outcome for each and all of the claimants or respondents separately.

Guillaume Zambrano also tested Meta’s Llama3 (70B) and Mixtral (8x7B) to extract structured data from French Courts of Appeal decisions.^[13] Specifically extracting the trial court outcome and the appellate court ruling, Zambrano utilised JavaScript Object Notation (JSON) prompting (example below). He concluded that case law extraction can be accomplished without the need for model fine-tuning or access to proprietary models.

Role: Lawyer.
Input: french judicial decisions.
Objective: Extract data.
Format: json UTF-8:

{
“decisionDate”: “YYYY-MM-DD”,
“decisionID”: “string”,
“conseilPrudhommesCity”: “string”,
“courtOfAppealCity”: “string”,

Legal Research

Research is a cornerstone of legal practice and traditionally requires significant time from professionals as they navigate primary and secondary sources. AI models can be helpful for research purposes, including extracting legal principles, summarising texts, identifying key concepts and brainstorming.^[14] However, they must be used as an aid rather than a replacement for traditional legal research methods, with appropriate verification and validation of sources. The University of Queensland et al. in Legal Research Skills: An Australian Law Guide set out some of the key risks and benefits of using generative AI for legal research.^[15]

Table 1 – Risk and Benefits of using generative AI for legal research
Potential Risks	Potential Benefits
Users need expertise to understand law and specific areas of the law in order to develop good prompts, ask effective questions and evaluate the genAI output for accuracy.	GenAI tools may suggest relevant case law that was missed during manual legal research.
Propensity of gen AI tools to hallucinate and produce inaccurate, non-factual information.	GenAI tools may flag discrepancies between jurisdictions that might initially be missed.
Hallucinations include citing case law and legislation inaccurately.	Specific search skills, such as applying Boolean techniques, are not needed in genAI tools.
Output may be unrelated to the jurisdiction in question and not give enough depth or analysis for the specific legal topic.	Iterative prompting facilitates deeper diving into legal topic areas.
Presence of bias or mis-/disinformation in genAI outputs due to the generic, global nature of information it draws from and/or malicious actors.	GenAI tools can summarise large amounts of information to distil the critical matters faster.
Negligent use or upload of proprietary, confidential or copyrighted materials to gen AI tools without applying professional judgement or the necessary genAI tool settings to meet professional and client obligations.
Possible negative impact on human capability to ideate, analyse, interpret and synthesise information.
Lack of user awareness of how AI and algorithms work leading to over reliance on genAI output.

Table 1. Risks and benefits of using generative AI for legal research by The University of Queensland et al. licensed under CC BY-NC

The authors offer the following prompt patterns for legal research purposes:^[16]

Act as a legal practitioner who is advising a client. Conduct legal research on [legal issue or topic]. Summarise the relevant case law, statutes and delegated legislation in [jurisdiction].

Provide analysis and conclusions based on your research. Please be concise and accurate.

Do not hallucinate or respond creatively.

Act as a legal practitioner who is advising a client. Provide an overview of the legal precedents in [area of law].

List the leading cases that exist in this area of law and the key legal arguments for each case. Provide analysis and conclusions based on your research.

Please be concise and accurate. Do not hallucinate or respond creatively.

Act as a legal practitioner who is preparing a case. Which are the governing statutes and delegated legislation relating to [legal issue] in [jurisdiction]?

How do these statutes and delegated legislation compare to statutes and delegated legislation from other Australian States and Territories.

List the leading cases that have cited this legislation and the key legal arguments for each case.

Please be concise and accurate. Do not hallucinate or respond creatively.

However, it is essential to note that prompts like these assume the model has been trained on relevant legal information or operates an RAG that contains legal information. The following AI tools have been developed for the Australian legal context:^[17]

Lexis Nexis Lexis+AI – https://www.lexisnexis.com/en-us/products/lexis-plus-ai.page (paid model)
Thomson Reuters CoCounsel – https://www.thomsonreuters.com.au/en-au/products/cocounsel.html (paid model)
ChatGPT Australia Law – https://chatgpt.com/g/g-gEBNOoaEJ-australian-law(limited free use)
Court Aid – https://courtaid.ai/(limited free use)
Legal Assistant – https://legalassistant.au/(limited free use)
Legal Genie – https://legalgenie.com.au/(free)

For specific research tasks, prompting chaining, combined with chain-of-thought and generated knowledge prompting, can be beneficial. Structure your prompts to guide the model towards finding relevant general information, starting broad and then narrowing down. For example:

Step 1

You are an Australian legal assistant working in intellectual property law. Provide an overview of copyright law in Australia.

Step 2

Now, focus specifically on fair dealing exceptions.

Step 3

Finally, thinking step-by-step, show how fair dealing could be applied to digital sampling in music.

At the intersection of text extraction and legal research, David Colaursso presents the following prompt pattern to extract citations from scholarly articles and save them to a comma-separated values (CSV) file (a text file format used to store tabular data in plain text).^[18] This example can be modified for other purposes (e.g., collecting legislation and case law citations in judicial decisions).

I’m about to show you a selection of text from an academic paper, complete with citations to other works. Your job is to find all the citations and produce the text for a csv file containing a row for each citation. The csv should have the following headers: “author”, “title”, “name_of_publication”, and “date”.

Consequently, each row should have the following data all encased in double quotes: author, title, name of publication, and date. Be sure to add a carriage return/line break to the end of each row, including the header row. If part of the information isn’t available return “unknown” for that part.

TEXT FROM ACADEMIC PAPER

Now provide a line for the header and each citation (row) in the csv file. Return only the text of the file. Do NOT place it in quotes or provide any extraneous text, just the header and rows.

The following sections present prompt patterns designed to assist with various research tasks and augment RAG, ranging from case law analysis to statutory interpretation.

Analysis

Research by Yu, Quartey, and Schilder has examined how various prompting techniques can enhance the performance of language models in legal reasoning tasks.^[19] It focuses specifically on the COLIEE legal entailment competition, which is based on the Japanese Bar exam. The researchers evaluated different prompting strategies using GPT-3 on legal entailment tasks to ascertain whether a given legal hypothesis is true or false based on the provided statutory premises. The best results were achieved through a zero-shot approach employing legal reasoning prompts rooted in specific legal reasoning frameworks, particularly TRRAC (Thesis, Rule, Rule, Application, Conclusion). The paper explored nine distinct frameworks, presented in the following format:

Please analyse if the hypothesis is True or False according to the given legal reasoning approach.

Approach: [Specific legal reasoning framework]

Premise: [Legal statute]

Hypothesis: [Statement to evaluate]

True or False?

An example prompt from the paper included this one that used IRAC:

Please analyse if the hypothesis is True or False according to the given legal reasoning approach.

Approach: Issue, rule, application, conclusion.

Premise: Article 509 The obligor of either of the following obligations may not duly assert a set-off against the obligee; provided, however, that this does not apply if the obligee acquires a claim corresponding to the relevant obligation from another person: (i) an obligation for compensation for loss or damage based on a tort committed in bad faith; or (ii) an obligation for compensation for loss or damage for death or injury to person (excluding the one set forth in the preceding item).

Hypothesis: If a person that holds a monetary claim has borne an obligation for compensation for loss or damage for injury to the obligor of that monetary claim, the person may not duly assert against the obligor, a set-off using the monetary claim against the obligation for compensation.

True or False?

Anthropic provides the following complex prompt pattern for a legal use case.^[20]

You are an expert lawyer. Here is some research that’s been compiled. Use it to answer a legal question from the user.

When citing the legal research in your answer, please use brackets containing the search index ID, followed by a period. Put these at the end of the sentence that’s doing the citing. Examples of proper citation format:

Example 1: The statute of limitations expires after 10 years for crimes like this. [3].

Example 2: However, the protection does not apply when it has been specifically waived by both parties. [5].

Write a clear, concise answer to this question:

It should be no more than a couple of paragraphs. If possible, it should conclude with a single sentence directly answering the user’s question. However, if there is not sufficient information in the compiled research to produce such an answer, you may demur and write “”Sorry, I do not have sufficient information at hand to answer this question.””.

Before you answer, pull out the most relevant quotes from the research.

Legislation

A different application involves requesting a model to generate a decision tree or flowchart from legislative documents. Experiment with the prompt provided below.

Transform the following statutory provisions into a [decision tree]/[flowchart] structure that could guide compliance determinations. For this transformation:

Identify all conditional statements (if/then relationships)
Extract defined terms and their definitions
Map exception pathways and alternative compliance routes
Include statutory references at each decision point
Structure the logic to conclude with clear compliance outcomes

The output should be described in a format that could be easily converted to a visual flowchart by a designer, with clear instructions on the hierarchical relationships between elements.

Statutory text:

[paste statutory text]

Case Briefs and Summaries

David Colaursso provides the following prompt pattern for creating case briefs. The model is given the text of a judicial decision to follow.^[21] The prompt can create case briefs and explain cases in plain language. The model could then directly answer questions about the case using prompt chaining.

In a moment, I will give you the text of a legal opinion, also known as a case. I will ask you to pull out the important bits and use them to write a Case Brief. Keep the following in mind when writing your brief.

Guide to Writing a Case Brief for Law School

Purpose of a Case Brief: Case briefs are essential study tools in law school, aiding in the analysis and recall of complex legal materials. They serve as a concise summary after thoroughly reading, dissecting, and understanding a case. Primarily for personal use, they also act as quick references for class discussions and exam preparations.
Audience: Primarily, you are the audience for your case briefs. Professors value the process but typically won’t review your briefs. In professional practice, the focus is on effective legal representation rather than the briefing process itself.
Essential Elements: Your brief should include the following core components for it to be effective:

Facts: Case name, parties involved, factual background, procedural history, and the judgment.
Issues: The legal questions or disputes at the heart of the case.
Holding: The court’s legal decision or rule applied in the case.
Rationale: The reasoning behind the court’s decision.

Optional Elements: Depending on the case, consider adding:

Dicta: Important but non-decisive commentary.
Dissent: Valuable opposing opinions.
Party’s Arguments: Each side’s main arguments.
Comments: Personal insights or notes.

Organization Tips: You might find it helpful to further break down the Facts into:

Specific case facts
Procedural history
Judgment outcome

Remember, the Holding differs from the Judgment as it reflects the legal principle, whereas the Judgment is the case’s final decision.

Relevance and Brevity: Focus on including relevant information that aids memory and understanding:

Include crucial facts, especially those pivotal to the case’s outcome.
Highlight the main legal issue(s) and the court’s final ruling.
Summarize the rationale, focusing on key reasoning and factors influencing the decision.

Length: Aim for your briefs to be concise, ideally one page, to facilitate easy review and reference.

—

Okay, now here is the text of the case.

—

When you write your brief, be sure that it includes the following sections:

Case Name:
Court:
Date Published:
Citations:
Parties:
Facts:
Procedural History:
Issues:
Holding:
Rationale:

If asked any follow-up questions, use the above text of the case, and ONLY the above text, to answer them. If you can’t find an answer in the above text, politely decline to answer, explaining that you can’t find the information. You can, however, finish a thought you started above if asked to continue, but don’t write anything that isn’t supported by the above text. And keep all of your replies short! But first, give me your case brief for the above case.

Judicial Predictions

Legal judgment prediction (LJP) is an emerging field that leverages technologies, particularly LLMs, to forecast the outcomes of legal cases. The ethical dimension of LJP is paramount, as it raises questions about fairness, accountability, and the potential consequences of relying on algorithmic judgments in legal contexts. Researchers have investigated whether LLMs can perform LJP. Two research papers that have explored prompt engineering to facilitate LJP are discussed below.

Trautmann, Petrova, and Schilder tested prompt engineering to conduct LJP tasks.^[22] They investigated the performance of zero-shot prompting for specific facts in case texts from the European Court of Human Rights and the Federal Supreme Court of Switzerland. The prompt structure they utilised is illustrated in the figure below.

A diagram showing a five-step prompting pipeline for processing a long document with a large language model (LLM). — An example prompt template in English for an ECHR task by Trautmann, Petrova and Schilder is licensed under CC BY-NC-ND 4.0

They explained the approach as:^[23]

Step 1: We use just the long legal document (one at a time) as the only input to the model. The results were a continuation of the document with other potential facts since these are valid completions. However, this was not useful for our task.

Step 2: We included a question after the document, which is a reformulation of the text classification task. This improved the output of the model, but it was still not working for many cases, where the model continued with a list of other questions.

Step 3: The inclusion of the “Question:” and “Answer:” indicators was again improving the completion by the model. Unfortunately, this was also a “free-form” response and it was difficult to map to one of our two intended classes (yes / no).

Step 4: We included the answer options “A, Yes” and “B, No” and finally

Step 5: The special GPT-model indicator “<|endoftext|>”, to separate the document from the prompt.

In their paper, Cong Jiang and Xiaolei Yang propose legal syllogism prompting (LoT) to use LLMs for legal judgment prediction.^[24] Unlike chain-of-thought prompting, LoT is zero-shot without additional examples. A legal syllogism is a form of deductive logic.^[25] Think of the expression:^[26]

Major premise: All men are mortal.

Minor premise: Socrates is a man.

Conclusion: Therefore, Socrates is mortal.

The authors assert that ‘LoT can activate the model’s deductive reasoning ability, allowing the model to output the legal reasoning process and conclusion.’^[27] Although both papers sought to complete LJP tasks, the results were not very favourable. However, they illustrate approaches to prompt engineering that encourage models to engage in legal reasoning.

Semantic Searching

Emerging AI legal research tools, such as Habeas, are reimagining the way legal information is searched. Unlike traditional search techniques that rely on Boolean search strings and exact match keywords to find legal information, these tools utilise semantic search capabilities powered by advanced machine learning and natural language processing (NLP) technologies.

Semantic search represents a significant advancement in legal research capabilities, moving beyond traditional keyword matching to understand the meaning and context of legal queries. Unlike conventional Boolean search methods that rely on exact word matches, semantic search uses natural language processing to comprehend the intent behind research questions and retrieve relevant results based on conceptual similarity rather than mere keyword occurrence.

Understanding Semantic vs. Traditional Search

Traditional legal research databases require practitioners to master complex Boolean operators, field codes, and precise terminology. A typical search might look like:

"duty of care" /s negligence /p "reasonable person".

Semantic search allows practitioners to ask questions in natural language, such as: “What are the elements required to establish a duty of care in negligence cases involving economic loss?” The AI system interprets this query, understands the legal concepts involved, and retrieves relevant cases, statutes, and commentary even when they use different terminology or phrasing.

Effective Semantic Search Prompting

When using semantic search tools, consider these prompt patterns:

Conceptual Queries

Instead of rigid keyword combinations, frame searches around legal concepts:

What are the requirements for establishing unconscionable conduct under the Australian Consumer Law?

Rather than a traditional search such as: "unconscionable conduct" AND "Australian Consumer Law" AND elements

Scenario-Based Searches

Describe factual situations to find analogous cases:

Cases involving directors’ duties where a company entered into transactions with entities controlled by family members of directors.

Comparative Analysis Requests

Ask for analysis across multiple legal areas:

How do courts balance freedom of speech against anti-discrimination protections in employment cases?

Temporal and Jurisdictional Constraints

Include specific limitations in natural language:

Recent Federal Court decisions since 2020 on the application of the deceptive similarity assessment in the Trade Marks Act 1995 sections 44 and 120.

Using this search, Habeas came up with this summary of results:

LEGAL RESEARCH SUMMARY

Summary

Recent Federal Court decisions since 2020 have continued to apply well-established principles for assessing deceptive similarity under sections 44 and 120 of the Trade Marks Act 1995, with particular emphasis on the likelihood of confusion in the minds of consumers.

Detailed Analysis

Core Legislative Framework

Section 10 defines deceptive similarity as when a mark “so nearly resembles that other trade mark that it is likely to deceive or cause confusion”
Section 120(1) establishes infringement when a person uses a sign that is substantially identical with, or deceptively similar to, a registered trade mark

Key Principles for Assessment

Consumer Perspective

The test relates to consumer recollection of the mark in its trade context [1]
Courts must consider the “imperfect recollection” of consumers rather than direct side-by-side comparison [4]
Assessment must account for the “general impression” created by the marks [4]

Standard of Proof

Must demonstrate a “real tangible danger of deception or confusion” – mere possibility is insufficient [1],[5]
Confusion exists if consumers are caused to wonder whether products/services come from the same source [5]

Contextual Considerations

Courts must examine:

Circumstances of mark usage
How goods/services will be bought and sold
Character of probable purchasers
All surrounding circumstances [5]

Recent Application

In Caterpillar Inc v Puma SE [2021] FCA 1014 [3], the Federal Court reaffirmed these established principles while emphasising the importance of considering:

The essential features of both marks
The overall impression created
The context of actual marketplace use

Key Considerations

1. Assessment Limitations

The comparison is not based on direct visual comparison [2]
Must consider all goods/services within registration specification, not just current use [5]

2. Practical Application

Focus remains on real likelihood of confusion rather than theoretical possibility [1]
Assessment must consider actual marketplace conditions and consumer behaviour [5]

3. Timing

Rights are determined as at the date of application [5]
Must consider potential future use scenarios within registration scope [5]

REFERENCES

[1] Global Brand Marketing Inc v Cube Footwear Pty Ltd [2005] FCA 852