Every Holmes Needs a Moriarty
Intelligence goes soft without opposition. For the past year I’ve run two AI platforms against each other on everything that leaves my desk — one to draft, one with a single job: break the draft. Here’s the method, and where the adversary earns its keep versus where it cheats.
Conan Doyle understood something about intelligence that most of us learn the hard way: it goes soft without opposition. Holmes called Moriarty the Napoleon of crime and admitted that his horror at the man’s crimes was lost in admiration of his skill. Holmes was at his best when an intellectual equal was pushing him.
I think about that every time I finish a draft. For the past year I have been running two AI platforms against each other on nearly everything that leaves my desk with my name on it. One platform drafts. The second platform has one job: break the draft. I call the instruction I hand the second one an adversarial brief, and it has become the most reliable quality step in my process. What follows is one operator’s method, not a study. But it is a method I use every week, and it works.
Why one platform can’t check itself
The obvious move is to ask the platform that wrote the piece to critique it. Ask for ordinary feedback and the usual result is compliments with a few gentle suggestions attached. A model reviewing its own output brings the same training, the same habits, and the same blind spots that produced the draft. It can catch some of its own mistakes, but you are asking the same mind to find what the same mind missed.
A second platform brings different training, different tendencies, and different failure patterns. It will catch things the first one missed in its own work, not because it is smarter but because it is different. That is the Holmes-and-Moriarty dynamic.
Watson admired Holmes. Moriarty tested him. When quality is on the line, you want a tester.
The adversarial brief
The instrument that makes this work is the brief. If you paste a draft into a second platform and ask for feedback, you will get the AI equivalent of a polite golf clap. You have to assign it the adversary’s job explicitly.
My briefs do three things:
- State the mission. Its job is to find what is wrong with this document — claims that cannot be traced to a source, numbers that do not hold up, structure that does not serve the argument, and paragraphs that are not earning their place. If I want numbers or claims checked, I attach the source material, because a reviewer without the sources is guessing.
- Fence off what is settled. Every draft contains decisions I have already made and do not want relitigated: the audience, the length, the position the piece takes. I list those explicitly so the reviewer attacks the open questions instead of rearguing closed ones.
- Demand specifics with reasons. “This section is weak” is useless. “This number has no source and your reader will notice” is gold.
What the adversary gets right, and where it cheats
After a year of this, I can tell you where the second platform earns its keep and where it goes too far.
It earns its keep on accuracy and structure. It has caught unsupported statistics, a sign error in a financial calculation, and sections that repeated an argument I had already made somewhere else. Those catches alone justify the step. But a catch is a lead, not a verdict. The critic can be wrong too, so the flagged number still gets checked against the source before anything changes.
It is unreliable on voice. The adversarial reviewer’s consistent instinct is to sand precise language down into safe, general hedges. It votes for flat every time, and if you accept those edits, your work stops sounding like you and starts sounding like everything else. That is a Moriarty move, in its way. Give a model the critic’s job and it tends to keep finding things to criticize, even when the safer edit is not the better edit.
Take the accuracy findings seriously. Treat the voice recommendations with suspicion.
Which is why the final call always belongs to you. Holmes solves the case. Moriarty just made him sharp enough to do it.
How to run it this week
You do not need to be publishing columns for this to pay. A pricing memo, a customer letter, a policy draft — anything with consequences qualifies. The pass supplements the review you already do. It does not replace your accountant or your attorney.
- Draft on your primary platform the way you normally would.
- Open a second platform from a different provider. Not another tab on the first one, and not the model that drafted.
- Give it an adversarial brief. State that its job is to find flaws, list the decisions that are settled and off limits, attach the source material for anything you want fact-checked, and require specific findings with reasons rather than general impressions.
- Sort the findings. Verify accuracy catches against the source or the underlying math before accepting them. Judge the structure catches. Give the voice and tone suggestions the most scrutiny of all.
- Make the call yourself. The adversary informs the verdict. It does not render it.
At Reichenbach Falls, Holmes survived and Moriarty did not. Your draft should fare better. Keep the adversary working for you, let it test everything that matters, and reserve the last word for the one person in the room who actually owns the outcome.
