The Superstar Paralegal Skill You Need!
What you’ll learn:
Why AI’s biggest problem isn’t wrong answers, it’s confused half-finished rewrites dressed up as clean
The eight things I make it check before I believe any output is actually done
The exact skill, ready to drop into whatever tool you’re already using
AI acts like a human more than we give it credit for. Give it imprecise instructions, keep the conversation running long enough, and it gets confused the same way a person would. It loses the thread. It forgets what you told it three turns ago.
But there’s one way it’s worse than a human. A paralegal who catches an error fixes that error. AI, more often than not, rewrites the whole thing. Ask it to fix one clause and you get a document that’s been quietly touched everywhere, some of it better, some of it not, and now you have to reread all of it to find out which is which.
So I stopped trusting “looks good” from AI on anything that mattered. Not because it’s dumb. Because a confused, half-finished pass dressed up as a clean one is worse than no pass at all. I built a skill that forces it to check, category by category, and say exactly what it found, so I’m never taking “done” on faith.
The deal terms are settled. The partners signed off on substance. All that’s left between a polished document and an embarrassing one is the read nobody wants to do by hand. This is what I hand to AI instead, and what I make it prove before I believe it.
Not the law. The typos. The defined term that lost its capital letter three drafts ago. The cross-reference pointing at a section that got deleted two rounds back. The signature block someone copied from a different deal, address and all.
The eight things it checks
Typos and grammar, obviously. Misspellings, doubled words, “their” for “there,” subject-verb agreement. This is the category any tool already handles reasonably well on its own.
Defined terms. This is where most contracts actually fall apart. “Company” shows up lowercase somewhere it shouldn’t. A term gets used three pages before it’s defined. Two different definitions sneak in for the same word because nobody deleted the old one when the deal changed shape.
Numbering. Sections run straight through, no gaps, no duplicates, no Section 5.3 floating around with no 5.1 or 5.2 anywhere near it.
Cross-references. “As set out in Section 4.2” needs to actually point at a Section 4.2 that says what it claims to say. This is the tell for a document that got renumbered halfway through drafting and nobody went back to fix the references pointing at the old numbers.
Citations. Complete, formatted the same way every time it appears. And if something looks off, it gets flagged, not quietly patched. I don’t want it guessing at what a citation probably should say.
Dates and formatting. One date format for the whole document. Signature blocks that actually match the parties named in the body, not close enough.
The right party, named everywhere. This is the one that can actually sink a deal. Wrong entity, right entity with the wrong address, a signatory who isn’t authorized for that specific entity. Checked in the recitals, the defined term, the notices clause, the signature block. Every single place the party shows up, not just the first one.
Placeholders. The bracket nobody filled in. The “TBD” that made it into what was supposed to be the final version. Comments and tracked changes still sitting open in a document about to go out for signature.
The rule that matters more than the categories
If it’s not sure something is wrong, it says so instead of guessing either way.
That sounds small. It’s the whole thing. I’d rather get ten borderline flags I have to glance at and dismiss than have it quietly decide something is probably fine and say nothing. And I never want it inventing an answer to fill a gap. Not a citation, not an address, not a cross-reference target. If it can’t confirm something, it says unverified. That’s a different answer than clean, and it has to actually say which one it’s giving me.
That’s why I trust the pass. Not because it’s smart. Because it doesn’t pretend to know things it doesn’t.
How it reports back
Grouped by category, not one long list. For each issue: where it is, what’s wrong, what I’d do about it. If a category came back clean, it says so, so I know it was checked and not just skipped over.
Then two lines at the end. What it actually checked, in plain language. And a verdict, either clean or a count of what needs fixing, with an honest flag if anything came back unverified instead of folding it into a clean bill of health.
What it doesn’t do
It’s not judging the deal. It’s not flagging the placeholder someone left open on purpose because that term is still being negotiated. It’s not stripping tracked changes because they’re sitting there.
That’s a different job entirely. Keeping this one narrow, pure mechanics and consistency, is exactly what makes it worth trusting. The second you ask it to also weigh in on the substance, it gets worse at catching the wrong entity name.
The practical takeaway
Don’t take “done” from AI on faith, especially after a long back-and-forth where instructions piled up and got vaguer with each round. Run this, category by category, and make it tell you exactly what it checked.
Zero typos. Zero inconsistent terms. Zero broken cross-references. Zero citations nobody verified. Zero wrong-party mistakes. Zero forgotten brackets. That’s the bar. It’s not complicated. It just has to actually get checked every time, not assumed clean because the output looked confident.
Full skill file is below. Drop it into Claude, ChatGPT, Wordsmith, whatever you’re already using. Custom instruction, project file, system prompt, doesn’t matter. Same pass, every time.
Speak soon!