Most teams use AI backwards.
They start with the prompt. The model drafts, the team edits, the copy ships, and six weeks later somebody asks where the 40 percent number came from. Nobody remembers. The number was confident, so nobody checked.
There is a better order. It has four stages: a claims brief before anything, generative drafting inside strict boundaries, machine verification of every claim the draft makes, and a human editor who owns the final word. It is the process I run on every Kaperider engagement, with the Lucidry refinery as the verification stage, and this article walks through each part the way I would show a new client.
You do not need my software to run it. You need the order.
What comes before the prompt?
The claims brief.
Before any generation, write down every claim the finished piece is allowed to make, in three columns:
The claim. What you will assert, in plain words. "Our platform cuts onboarding time for enterprise customers."
The evidence. What backs it. A customer study, a benchmark, an internal number, a named example. If the evidence column is empty, the claim is not allowed in the copy, or it gets rewritten to what you can prove.
The protected fields. The names, numbers, quotes, and citations that must survive every edit unchanged. Product names, client names, dates, prices, legal lines.
A claims brief for one landing page usually fits on a single page. That page is the difference between generating copy and generating liability.
Why does it come first? Because the brief is the only stage where you decide what is true. Every stage after this one can move fast precisely because the truth was decided slowly, by a person, before a machine touched it.
The most common objection: it takes an hour and the model would have taken a minute. Correct. The brief is the hour that saves the quarter. One wrong claim in a sales asset costs more time than a hundred claims briefs.
Draft inside boundaries, not open prompts
The model gets the claims brief, the positioning, and the audience, and is asked to work inside them.
The model gets the claims brief, the positioning, and the audience, and is asked to draft inside them. The prompt carries the rules that matter for your situation: never introduce a number that is not in the brief, never invent a customer, never invent a study, never soften or harden a guarantee.
Then you ask for candidates, plural. Three openings, two structures, one alternative angle. Generation is cheap; choosing is expensive, so spend the machine's cheapness on options and your attention on selection.
What you do not do is prompt first and verify later. A model asked for "10 statistics about content marketing" will produce ten confident statistics, some real, some plausible inventions. The failure mode is not that the model refuses to write. It is that it writes too well to be doubted.
Draft with the brief open beside you. If a draft says something the brief does not cover, mark it and check it before it survives. The draft is a candidate, not a source of truth.
The machine catches the quiet failures
A person reading a draft checks whether it sounds good. Machine verification checks whether it is still true.
Source checks tie each claim in the draft back to the evidence column of the brief. If the draft asserts something the brief never sourced, it gets flagged. This catches the model's habit of upgrading "many customers" into "thousands of customers."
Contradiction detection compares the draft against the other assets your company has shipped. Your new landing page against your sales deck. Your email sequence against your pricing page. Claims drift apart slowly, one edit at a time, and drift is invisible up close. The Medium State of Writing survey found around 60 percent of writers do not trust their own AI-assisted work, and cross-asset drift is the reason most of them are right.
Protected-field integrity watches the names, numbers, quotes, and citations. When an edit changes "8,200 publishing professionals" to "8,000 publishing professionals" because it rounded for rhythm, the check catches it and puts it back.
In the Lucidry refinery this stage ends with a receipt: a change log of every edit, and an independent quality score produced by a scoring pass that never saw the draft's origin, so the score cannot flatter the process that produced it. The receipt is the deliverable's provenance. It ships with the copy.
What does the human actually do in human-in-the-loop editing?
The part that cannot be delegated.
The human editor makes the judgment calls the machine correctly refuses to make. Is this claim strong enough to carry the page? Is this the client's actual voice, or the model's average of every voice? Does the sequence land the argument in the right order, or does it just include all the pieces?
The editor accepts, rejects, and rewrites against the brief. When the refinery flags a contradiction, the editor decides which asset is right. When the score comes back lower than expected, the editor finds out why instead of shipping anyway.
And then the human does the one thing no model can do, which is to put a name on it.
Human-in-the-loop is not a euphemism for "a person glanced at it." It is a specific commitment: a named professional examined every claim, decided what survives, and is accountable for the result. That is the difference between AI-assisted copy and AI-abandoned copy. The loop closes with a signature or it does not close at all.
You can run this without Lucidry
Yes, degraded, but yes.
The claims brief is a document you can write today. The drafting boundaries are prompt discipline. Verification can be manual: a checklist where each claim gets checked against its source, each number against its original, each asset against the master blueprint. It is slower, it scales worse, and it depends on the same person who is tired at 5 p.m.
That is why the refinery exists. Not because the process is a secret, but because doing it by hand at volume does not survive contact with a real content calendar. The machine checks at machine speed so the human can spend attention where it counts.
Steal the order even if you never buy the tool. Brief. Draft. Verify. Sign.
How much longer does this take?
For the first piece, about twice as long as prompting raw.
For every piece after the first, less. The brief is reusable. The prompt rules are reusable. The blueprint holds so the contradictions stop being re-found every quarter. And the receipt replaces the quiet dread of not knowing what your own website claims.
The alternative is faster on day one and unfixable on day one hundred. Teams do not lose their facts in one bad draft. They lose them in a hundred fast ones.
Start with the asset that worries you most
Not the one you need next. The one you quietly avoid opening. Write the claims brief for it, one page. Then run your normal generation, and check the draft against the brief.
The gap between those two documents is your actual content risk, in claims, and now you can see it.
If you want the process run for you, that is the work: the messaging blueprint, the proof of work, and The Offer that shows what an engagement includes. The two-minute brief form starts it.
Kaperider is the copywriting practice. Lucidry is the refinery behind the work.