AI prompts that are tested, structured and built for the real day-to-day work of lawyers and legal professionals. Not generic one-liners — working tools.
Each prompt comes with a 🔌 MCP Factory variant for maximum reliability
A prompt is the instruction you give a generative AI to obtain a result: a question, a directive, a full brief with attachments. French has had an official term for it since September 2024, the Commission d’enrichissement de la langue française having settled on “instruction générative” (generative instruction): the directive, usually written in plain language, that a user addresses to a generative model to describe the task expected of it. In law firms as everywhere else, everyone still says prompt.
The official definition holds the essential point: a prompt describes the task. A good instruction therefore goes beyond a plain question and brings together up to five elements:
who the model is to play (“employment law associate, litigation before the conseil de prud’hommes, the French labor tribunal”)
the facts, the file, what the work is for
precisely what you expect
a table, a two-part outline, a checklist, a draft clause
permitted sources, governing law, tone, length
Why this much precision? A large language model (LLM) generates text by predicting the most likely continuation of your request. It has no access to your file, to your strategy, or to the state of the law on the day you ask, unless you supply them.
Every piece of information missing from the prompt is replaced by a statistical guess. In law, a statistical guess quickly becomes an error.
Dismissal analysis, severance calculation, labor tribunal arguments, HR compliance, case law monitoring.
Commercial leases, transactions, co-ownership, urban planning, construction, real estate litigation.
Due diligence, SPA, reps & warranties, earn-out, locked-box, shareholders' agreement, LBO, closing memo.
Legal characterization, police custody, nullities, pleadings, sentence enforcement, white-collar crime.
Divorce, parental authority, child support, compensatory allowance, protection of vulnerable adults.
The techniques that made the difference in 2023 have become useless, sometimes harmful. Here is what moved, and what holds.
ChatGPT has been public since late November 2022 and prompt engineering becomes a discipline in its own right, with its recipes — “act as an expert”, “think step by step” (chain-of-thought), two or three examples to frame the output (few-shot). Companies hire prompt engineers on six-figure salaries. Those techniques genuinely changed the output, because the models of the day were brittle and every word counted.
Context windows go from a few pages to several hundred. RAG (retrieval-augmented generation, answering from a document search rather than from memory) becomes standard, and instead of relying on what the model remembers you plug it into sources you trust.
Stanford study “Hallucination-Free?” (2024). For legal work the shift is measurable — grounding answers in documents reduces error without eliminating it.
Reasoning models plan their answer before writing it. The instruction “think step by step” becomes pointless, sometimes counter-productive, because it interferes with that internal reasoning. Best practice flips — describe the objective, the constraints and the context, then let the model pick its own method. In June 2025 Andrej Karpathy (formerly of OpenAI and Tesla) popularizes the term context engineering, the art of filling the context window with the right information at the right time. Attention moves from the wording to everything around it, from documents to tools and memory.
The prompt typed on the fly into a chat window loses ground to system prompts, to prompt libraries shared across a firm, and to agents wired to tools, notably through the MCP protocol. Prompt engineering does not disappear for all that, since every agent, every tool and every step of a system rests on written instructions. The two disciplines stack. Another lesson of the period, documented by Anthropic under the name context rot, is that past a certain volume each additional piece of information dilutes the attention of the model. A short, dense context beats an exhaustive one.
Three years of hindsight are enough to sort the lasting from the passing. Formulas and tricks have a shelf life of 12 to 18 months. Four fundamentals survive every generation of models:
Every prompt assigns the AI a specific legal role
The domain parameters are built in, not left to chance
Explicit instructions to limit made-up answers
Tables, structured outlines, checklists — no vague prose
These prompts are an entry point. To truly understand how LLMs work in legal practice, explore our tracks on Zevra Campus or our full study on AI training for lawyers in France (11 bar schools analyzed, AI Act, training landscape).
Discover our training programs to master AI in your legal practice, or get in touch for tailored support.
Zevra Campus Contact us