Here’s the most common mistake I see lawyers make with AI, including good, careful ones. Claude hands them a draft. Something is off: the format, a section they always include, how they phrase a provision. They fix it themselves, decide it’s good enough, and move on.
Next week, same draft, same problem, same fix.
That correction was thrown away. Your AI never learned it, because you never told it.
One more sentence
The fix costs you one sentence. When the output isn’t how you want it, don’t just edit it. Tell your AI:
I don’t like it that way. I want you to do it this way in the future. Don’t ever do it the other way again, and don’t let my staff do it the other way either.
It may ask you a few questions to clarify. Answer them. Then it updates the rule, the template, or the skill so the next draft starts where your correction left off.
Why this beats training a person
I’ve trained paralegals for years. You show someone how you want something done, and the next time it’s close but not quite. Then they go on vacation, or leave, and you start over.
Once my AI is taught something, it’s about 95 percent reliable at doing it that way going forward. It doesn’t forget. It doesn’t quit. And it applies the rule to everyone in the firm who uses it, so your standard stops depending on who happened to be trained.
That’s the real promise. Not that AI drafts perfectly out of the box. That your standards finally stick.
Little preferences count too
Everyone has stylistic preferences. I have mine, and one of my coaching clients has about ten she won’t compromise on. Those are exactly the things to teach, because they’re what make a draft yours instead of generic.
Format issues, section order, how holidays appear in a parenting plan, whether values show up on an equitable distribution chart, how your signature block looks. Each one is a sentence. Teach them as they come up, and the drafts get closer to final every week.
The edited draft is the best teacher
The strongest move is to give your AI your finished version. After you’ve edited a draft into something you’d actually send, hand it back:
Here is my final version after my edits. Look at my changes and remember how I did this for the future.
Now it’s not learning from a description of what you want. It’s learning from the real thing.
The same coaching client turned around a 50-page prenuptial agreement in 48 hours, built from a Zoom transcript, a pile of emails from her client and his CFO, and her own favorite prenup templates. About 45 minutes of her review and it went to the client as a draft. She still edited it, because she edits everything. But the base was hers: her templates, her structure, her standards.
Give it a home
Rules need somewhere to live. Two habits help.
One project per client. Open a new AI project for every client and do all of that client’s work inside it. You stop scattering work across dozens of chats, and “this case” always means the right case. My client asked for “a quick mediation summary for this case’s mediation” before a 10 AM session, and it came back in about 45 seconds because the project already knew the case.
Project instructions. Every project should have written instructions that tell the AI what the project is for, where files live, and the rules to follow. They keep it on task and stop it from drifting. If you don’t know what to write, ask your AI to write your project instructions, then paste them in.
Your rules, your system
Everything you teach this way is yours. It lives in your firm’s files, branded for your firm, and it isn’t locked inside a vendor’s product. I wrote about why owning your setup matters. This is how that ownership builds up over time, one correction at a time.
If you want help turning the fixes you keep making into firm rules, bring a draft you’re tired of correcting to Open Office Hours, Fridays 2:00 to 3:00 PM ET. Join on Zoom. Or book a custom build strategy session.
