Legal prompting: practical techniques to make generative AI work well in a law firm

Two lawyers use the same generative AI tool for the same task and get radically different results. The difference is not in the model but in the instruction. Prompting, the craft of framing the request well, has become another professional skill, and in legal work it has rules of its own. This guide collects the techniques that work best in a firm's daily practice.

Ground rule: AI does not know law, it predicts text

Before any technique, the warning that frames everything else: generative models do not query legal databases or "know" statutes; they predict plausible text. They can invent articles, case law and even complete judgments with case numbers and dates, a real risk that has already led to sanctions for AI-generated fake case law. Everything the tool produces is a draft that requires verification, never a source.

Technique 1: assign role, task and format

The basic structure of a good legal prompt has three parts:

• Role: "Act as a Spanish employment lawyer experienced in collective bargaining."
• Task: specific and bounded: "draft a wage review clause for a company-level collective agreement", not "make me a collective agreement".
• Format: "in a single paragraph, in contractual language, without explanations" or "in a table with four scenarios".

The role sets the register, the task avoids generic answers and the format saves restructuring work.

Technique 2: give context without giving data

Models improve substantially with context: jurisdiction, sector, contract type, the position of the party you represent. But context has a clear boundary: client data. Before pasting anything into an external tool it must be anonymized, because professional secrecy applies to AI too. The practical rule: names, specific amounts and identifying facts are replaced by variables ("the employee", "amount X") and reintroduced in the final document.

Technique 3: iterate instead of aiming for perfection

The first result is rarely the good one. The efficient flow is conversational:

1. Ask for the draft with role, task and format.
2. Correct the approach: "more protective of the landlord's interests", "remove the case-law reference", "cut it in half".
3. Ask for self-critique: "point out the weaknesses of this text from the opposing party's position".
4. Ask for alternatives: "give me three versions of this clause, from most to least aggressive".

Cross self-critique is especially useful: asking the model to attack its own text brings out objections the initial draft was hiding.

Technique 4: keep a library of your own prompt templates

Firms that get value from these tools do not improvise: they maintain a library of tested prompts for recurring tasks (summarizing claims, comparative tables of legislation, first versions of pleadings, difficult emails). The template is refined with use and shared. It is the same principle that guides any internal use policy, like the one we developed analysing the AEPD's generative AI policy as a model for firms.

Technique 5: separate what AI does well from what it must not touch

An honest map of uses:

• Works well: first drafts, summaries of documents supplied by the user, working translations, checklists, argument outlines, client explanations.
• Works under supervision: statute searches (every citation must be confirmed in the official gazette), case-law analysis (only with verified sources), calculations.
• Must not be used: strategy decisions, substantive assessments that are not reviewed, any content that leaves the firm without a lawyer reading it in full.

Verification as the final implicit prompt

The CCBE guide and the Spanish Bar's charter agree on the essentials: final responsibility always rests with the professional. That makes verification part of the flow, not an extra. Three minimum checks before using any output: the cited statutes exist and remain in force (consolidated text in the BOE), the cited judgments exist (CENDOJ or the corresponding official database) and the facts match the file. The review of pleadings already has specific professional-conduct guidance on verification.

Those who master these techniques do not work less: they work on something else. AI removes the blank page; judgment, strategy and signature remain the lawyer's.