How lawyers can write better AI instructions for legal tasks

Why Workflow Control Matters

The role is not decorative language such as “act as an excellent lawyer.” It should identify the legally relevant perspective. In a supply-agreement review, the instruction might request analysis from the buyer’s counsel’s perspective. That directs attention towards the supplier’s obligations, acceptance mechanisms, liability, and termination. If several parties or interests are involved, the instruction should name the represented client. This prevents the result from mixing neutral observations with recommendations designed for the wrong contractual position.

Matter context should contain only the facts that change the analysis. A lawyer might state that the document is the supplier’s draft, negotiations are in their first round, and the service is operationally important to the client. The task should then use an observable action: identify departures from preferred positions, compare two versions, or prepare questions for the client. “Tell me what you think” offers no completion criterion. A precise task lets the lawyer determine whether every requested part of the assignment has actually been addressed.

A Familiar Legal-Work Scenario

Constraints matter when the documents do not contain everything needed. The instruction can require missing information to be flagged rather than guessed, dates and names to be preserved exactly, or the analysis to remain within the stated law and materials. Sources should be named explicitly: the current agreement, an approved internal template, and a selected negotiation note. If those materials do not support a conclusion, the output should treat it as an open question. This separates plausible prose from a legal work product that can be checked.

The output format should be chosen before generation begins. For contract review, a long narrative may be less useful than a table containing the clause, current position, identified departure, practical consequence, and proposed action. Meeting preparation may call for a question list, while a dispute file may require a chronology. Format is not merely a visual preference. It determines which information is separated, which relationships become visible, and how quickly the lawyer can test the result against the assignment.

Where Wisanna Fits

Compare two instructions. The vague version is: “Review this contract and tell me the problems.” A more useful version is: “Examine the draft supply agreement from the buyer’s perspective. Use only the current draft, the approved internal template, and the selected negotiation note. Identify departures concerning service levels, limitation of liability, termination, and personal-data processing. Do not assume facts absent from the documents; flag missing information. Present a table with the clause, departure, client impact, proposed response, and question requiring confirmation.” The second instruction gives both the AI and the lawyer a concrete standard for the result.

Wisanna is a private and secure legal-AI workspace designed for lawyers and professional legal work. In AI Chat, a lawyer can frame this instruction around the relevant documents and request a review matrix instead of generic commentary. The Microsoft Word add-in can support work where the document is being reviewed, while Wisanna Draft can turn the result into an editable legal document. These product surfaces do not make any output automatically correct or final. Their practical value depends on a well-specified assignment and professional legal review.

An end-to-end workflow can begin by selecting relevant prior material: an approved contract template and a negotiation note containing the client’s positions. The lawyer can apply that material to the current legal task by asking for a comparison with the new draft and a table of departures. Next comes checking each observation against the documents, correcting missing context, and deciding whether the proposal should be approved for the negotiation round or reused in a future instruction. Prior work is therefore not copied mechanically; it informs a current legal decision.

A Practical Evaluation Test

Wisanna can also support refinement before the substantive task begins. If an instruction requests analysis without identifying the represented party or the required form, the lawyer can rewrite it around those omissions. Within a team, the same structure can become a working habit: recurring assignments specify perspective, permitted documents, constraints, and deliverable format. Colleagues can then assess not only the answer, but also the quality of the instruction that produced it. This makes instruction design visible as part of legal work rather than an improvised exchange with a chatbot.

The final test is straightforward: could a colleague read the instruction and understand what must be completed, on the basis of which materials, and in what form? If not, the instruction needs another pass. A lawyer does not need technical language or an exceptionally long prompt. The goal is to describe the assignment with the same discipline used in good professional delegation. Role, matter context, task, constraints, sources, and output format turn a vague request into an AI-assisted legal assignment that can be reviewed, corrected, and used.