Part Two · The Doing · Chapter 9

The First-Draft Engine

There is a particular kind of tired that comes from writing the same thing for the hundredth time.

You know the documents I mean. The demand letter. The engagement letter. The routine response you have sent so many times you could nearly recite it in your sleep. The page looks blank when you start, but it isn't, not really, because you already know almost every word that belongs on it. Nine-tenths of it is the same as the last one. Only a few pieces change: the client's name, what happened to them, the number at the bottom. And yet every time, you build the whole thing again from the floor up, pouring good hours into typing words you have typed a hundred times before.

This is the third tool you will build, and it gives those hours back. Your model is going to write the first draft for you. Not the final draft. The first one. The difference between those two words is the whole chapter, so hold onto it.

The shape of a safe drafting tool

Think about what one of these documents really is. It is mostly fixed, with a few moving parts. The bulk of it is settled language: the phrasing you have refined over years, the paragraphs that must read exactly the way they read, word for word, because you chose those words on purpose. Then there are the few spots that change from case to case, the facts of this client, this injury, this demand.

That shape is the secret to building this safely, and it comes straight from the lesson of your first build. Remember the rule: never let the guesser be the part that has to be right. Your settled legal language has to be right. So you do not let the AI write it. Those words are yours, earned over years of practice, and they stay locked exactly as you wrote them. The AI never rewrites them. All you let it do is fill in the few blanks, the situational parts, using the facts you give it. The skeleton is yours and fixed. The AI only fills the gaps.

This is what people mean by keeping a human in the loop, except it is stronger than that. You are not in the loop. You are the loop. The AI is the thing in the loop with you.

Building your engine

Your engine is two pieces: a template and an instruction. And here is the part that will make you smile, because it means you barely have to do the boring work at all. You do not have to build the template by hand. The AI can build it for you, out of a letter you already wrote.

Start with your latest demand letter, a real one, a good one you were proud of. You are going to do this next part inside GPT4All, on your own machine, not on some website, because that letter is full of a real client's facts, and those facts stay home. That is the whole reason you have a private model. This is the moment you use it.

Open your letter in GPT4All and give it an instruction like this:

Here is a demand letter I wrote. Turn it into a reusable template. Keep all of my legal language exactly as it is, word for word. Find only the parts that change from one case to the next, the names, the dates, the description of what happened, the demand amount and its basis, and replace each one with a clear label in brackets, like [CLIENT NAME]. Do not change anything else. Do not invent anything.

In a moment it hands your letter back as a template, with the moving parts marked and your settled language left alone. It should look something like this:

Dear [OPPOSING COUNSEL NAME],

This firm represents [CLIENT NAME] in connection with injuries sustained on [DATE OF INCIDENT]. [ONE PARAGRAPH DESCRIBING WHAT HAPPENED, DRAWN FROM THE FACTS BELOW.]

[Your settled demand language here, fixed and unchanged.]

Accordingly, we demand [DEMAND AMOUNT] to resolve this matter. [ONE SENTENCE ON THE BASIS FOR THE DEMAND.]

Now check it, because this is the step you never skip. Read the whole thing. Did it bracket every part that changes from case to case? Did it leave a real client's name or a real number sitting in the text by mistake? Did it bracket a piece of your settled language that should have stayed fixed? Fix anything it got wrong, and when it is clean, save it. That is your template, and you will reuse it for years. You made it in about two minutes, and you typed almost none of it.

That is the first piece. The second is the instruction you give the model every time you want a new draft. Save this one too, so it is always ready:

Here is my template and the facts of a case. Fill in only the parts in brackets, using only the facts I give you. Do not change any other word of the template. Do not add anything I did not ask for. If you are missing a fact you need, write [NEED INPUT] in that spot instead of guessing. Keep the tone plain and professional.

Now you paste in your template, paste in the facts of the new case, and let it run. What comes back is your document, your settled language untouched, the blanks filled from the facts, and anything the model wasn't sure about flagged for you in plain sight.

That is the whole engine. A skeleton the AI helped you build and you checked, an instruction that keeps the AI in its lane, and a model that does the typing.

The line you never cross

Here is the rule that makes this tool safe instead of sorry, and it is the simplest one in the book. You never send what it gives you without reading every word.

Not skim. Read. The model filled in the facts, and the model can get a fact wrong. It can soften a sentence you needed hard. It can pick up the wrong number from your notes. It can write something that reads beautifully and is quietly not what you meant. None of that is a disaster, because it is a first draft, and a first draft is supposed to be wrong in places. That is what first drafts are for. Your job, the job that is still entirely yours, is to be the editor who catches it.

This is why the tool is a first-draft engine and not a finished-document machine. The name is a promise about who is still responsible. The AI gave you a running start. It did not give you permission to stop paying attention. Every word that leaves your office over your signature is yours, exactly as it always was.

What you built, and what you learned

Step back and look at the three tools you now have, because together they tell the whole story of doing this well.

The Deadline Keeper showed you when to keep AI out completely, because some things must be exact. Ask Your File showed you how to let AI read for you while you check every answer against the source. The First-Draft Engine shows you how to let AI write for you while you keep the settled language locked and stay the editor of every word. No AI, AI that reads, AI that writes. In all three, the same hand stays on the wheel. Yours.

That is not a small thing. Most lawyers using AI right now are doing it the dangerous way, pasting client secrets into a stranger's website and trusting whatever comes back. You are doing it the way that keeps your data home, keeps the exact things exact, and keeps you the lawyer. You built that. It runs in your office. It is yours.

In the next chapter we tie the pieces together, so that your private model and the tools around it stop feeling like three separate gadgets and start feeling like one quiet, private system that works the way you do.