A first draft from your own precedents

A model asked to draft a clause will produce something that reads like a clause. It will be generic, plausible, and unmoored from how your firm actually drafts — because it is drawing on everything it has ever seen rather than on what you have agreed to use.

The fix is not a better prompt. It is giving the model your precedent and constraining it to work from that.

This takes about forty-five minutes to set up for one document type, and it pays back from roughly the third use.

What you need first

Use only a tool, account and configuration your firm has approved for that material. An enterprise label is not the control: the decision still depends on the provider terms, retention, access, training use and the firm’s own confidentiality duties. The SRA’s guidance on client confidentiality is the regulatory starting point for SRA-regulated work; other jurisdictions have their own rules.

You also need to have chosen the precedent. Not a folder of them. One, the one you would actually start from, marked up so you know which parts are standard and which are negotiated. If you cannot name the precedent you would start from, the exercise will not work, and that is worth knowing on its own.

The instruction structure

Four parts, in this order. The order matters — later parts are read in the context of earlier ones.

1. The role and the boundary. State what is being produced and, more importantly, what must not be. “Produce a first draft of [clause type] for [transaction type]. Use only the structure and language patterns in the precedent supplied. Do not import standard wording from elsewhere.”

That last sentence does more work than anything else in the instruction.

2. The precedent, marked. Paste it, and label the parts. “Clauses 1–4 are standard and should carry through substantially unchanged. Clause 5 is negotiated and should be flagged for review rather than drafted.”

3. The facts. The specifics for this matter, listed. Where a fact is missing, say so explicitly — “the notice period has not been agreed” — rather than leaving a gap. A gap gets filled with an invention. A stated absence gets flagged.

4. The output shape. How you want it back. If you want a table of provisions with a column for what changed and why, ask for that. If you want prose, ask for prose. Unstated, you will get a different shape each time, which is what makes the whole thing feel unreliable.

The step people skip

Run it across twenty documents before you trust it.

Not one. One good result tells you nothing — it tells you the instruction can produce a good result, not that it will. What you are looking for is whether the same instruction produces the same structure across a realistic spread of inputs, including the awkward ones: the short-form version, the one with the unusual counterparty, the one where a clause is missing entirely.

Read all twenty. This is tedious and it is the only thing that converts an impression into knowledge. Record the actual distribution: how many passed, how many drifted in a consistent way, and how many failed for input-specific reasons. Do not start with an expected success rate. The measured result is the number you can quote to a supervising partner, and it is worth more than an adjective.

What this does not give you

It does not give you a draft you send. It gives you a first pass that starts from your own material instead of from nothing, which is a different and smaller claim.

Specifically it will not tell you when a provision is missing that ought to be there — absence is the hardest thing for these systems, because there is no text to react to. It will not tell you the commercial deal has changed shape. And it will produce confident, well-formatted output whether or not the precedent you gave it was the right one, which is why the choice of precedent is the part that still requires you.

Where the time actually goes

The first setup is forty-five minutes, most of it spent marking up the precedent rather than writing the instruction. The instruction is maybe ten minutes.

That ratio is the point, and it is the thing that transfers: the value is in having decided, explicitly, which parts of your own document are standard and which are negotiated. Most drafting knowledge of that kind is held tacitly by whoever has done it most. Writing it down is useful even if you never run the instruction again.