Prompt Engineering for Lawyers: Patterns That Actually Hold Up

Most prompting advice aimed at lawyers is a list of clever one-off tricks: a magic phrase, a role to assign the model, a format to request. These work, briefly, on the specific example the advice was written around, and then fail to generalize the moment a real matter's messiness shows up. What actually holds up is not a trick, it is a pattern, a reusable structure you can apply to a new task without reinventing it each time. Here are the five I actually use.

Pattern one: bound the pool before you ask the question

The single highest-leverage prompting habit, and the one most lawyers skip because it takes an extra minute of setup: never let a general-purpose model reason freely across everything it might know. Point it explicitly at the actual source material, the filed contract, the specific case file, the real precedent bundle, and instruct it to answer only from what is provided. This is the same discipline behind evidence anchoring, applied at the prompt level: a bounded pool is the single biggest reduction in fabrication risk, because the model has less room to reach for something plausible instead of something real.

The pattern: "Using only the attached documents, answer the following. If the answer is not supported by these documents, say so explicitly rather than inferring."

Pattern two: require the citation inline, not as an afterthought

A bibliography at the end of a memo is not the same discipline as an inline citation at the point a claim is made. The pattern is to build the citation requirement into the instruction itself, not to hope the model adds one voluntarily.

The pattern: "For every factual claim, cite the specific document, page and paragraph immediately after the claim. Do not state a fact without a citation attached to it."

Pattern three: build the "no answer found" path explicitly

Models under pressure to produce an answer will often produce one even when the honest answer is that the material does not support a conclusion. The fix is a prompt that makes silence an acceptable, expected output rather than a failure state, so the model is not implicitly rewarded for always having something to say.

The pattern: "If the provided material does not contain enough information to answer confidently, respond with 'not supported by the available documents' rather than guessing or extrapolating."

Pattern four: separate drafting from reviewing, even within one prompt

Asking a model to draft and simultaneously self-critique its own draft in a single pass tends to produce a weaker version of both tasks. The pattern that holds up is running them as two distinct passes, even if it is the same conversation: draft first, then a second, separate instruction asking specifically for weaknesses, unsupported claims or gaps in the draft just produced.

The pattern: first turn, "Draft [the document] from the following facts." Second turn, after the draft returns, "Review the draft above specifically for unsupported factual claims and gaps against the source material. Do not comment on style."

Pattern five: ask for the counterargument before you trust the conclusion

A model asked a direct question tends to answer it directly and confidently, which is exactly the failure mode that matters most in legal analysis: confident and wrong looks identical to confident and right until someone checks. The pattern is to explicitly request the strongest opposing position as a separate step, which forces the model to actually engage with the weaknesses of its own first answer rather than defend it by default.

The pattern: "Now argue the strongest position against the conclusion you just reached, using only the same source material."

Why these five and not a longer list

Every prompting trick I have seen that actually holds up across matters, not just the one example it was demonstrated on, reduces to one of these five moves: bound the input, force the citation, allow silence, separate drafting from review, and demand the counterargument. A longer list of tips tends to be five patterns wearing different clothes for different document types. Learning the patterns once is worth more than memorizing a hundred specific prompts, because the patterns transfer to a document type you have never prompted for before, and a memorized specific prompt does not.

The full set of prompt scaffolds built specifically for evidence-anchored legal drafting and review, not general prompting advice, is in the AI-Native Practice Field Guide.