AI Prompt Checklist for Legal Drafting

Ask a model to "draft a legal notice" and it will produce something confident, well-formatted, and quietly wrong. It will guess the jurisdiction, invent a limitation period, and assume facts nobody gave it. The draft looks finished, which is the danger: a partner skimming it may not notice that the cause of action was fabricated. A good prompt closes those gaps before the model starts, so the first draft lands close to usable instead of plausibly misleading.

Fix the document type and its reader

Name the instrument precisely. A demand notice under Section 138 of the Negotiable Instruments Act reads nothing like a reply to a legal notice or a covenant in a lease. State who receives it, whether that reader is hostile, neutral, or a client, and the register you want, formal court language or a plain explanation the business team can act on. Ambiguity here forces the model to average across styles and produce something that fits no one.

Supply the facts and mark what is unconfirmed

Paste the facts the draft may rely on, and label everything else. The model cannot tell a confirmed payment date from one you are still chasing, so tell it. A useful instruction is: use only the facts below; where a date, amount, or party detail is missing, insert a bracketed placeholder rather than a guess. This single line prevents the most common failure, a fluent sentence built on an invented particular.

State the governing law and forum

Indian drafting turns on specifics an untethered model will not assume correctly:

• The statute and section you are proceeding under
• The court or tribunal and its territorial jurisdiction
• Any contractual arbitration or exclusive-jurisdiction clause
• The limitation period, which you should verify yourself rather than trust

Give the model the framework; do not let it supply one. When it cites a provision, treat that citation as a lead to check against the bare act, not as authority.

Ask for structure, not prose

Tell the model the skeleton you want, recitals, then obligations, then notice period, then consequences of default, and ask it to keep each in a separate paragraph. A structured draft is far faster to review because you can scan for the clause that matters rather than reading a wall of text. For repeat instruments, keep a house structure and feed it in every time so drafts arrive in a shape your team already reads efficiently.

Set the boundaries of invention

The most valuable instruction tells the model what not to do: do not add clauses I did not ask for, do not assert a legal conclusion without flagging it, and do not smooth over a gap with a general statement. A model rewarded for sounding complete will paper over missing information unless you forbid it.

Review against the prompt, then the source

When the draft returns, read it against your own instruction first. Did it stay inside the supplied facts? Did it leave placeholders open, or fill them silently? Then check the substance against the underlying documents and the current statute. A tool like Juniorlawyer keeps the draft beside its source material so this comparison takes minutes, but the judgement stays with the lawyer who signs it.

A prompt is not a magic phrase. It is the same briefing you would give a competent junior: here is the document, here is the reader, here are the facts, here is the law, and here is what you must not assume. Brief the model that well and the first draft saves you real time. Brief it loosely and you inherit its guesses.