It started with a simple question. Two people use the same AI subscription — one wraps up a proposal draft in 40 minutes, the other spends three hours and still can't shake the feeling that something's off. Same access, same tools, yet the quality and speed of the output diverge. We set out to trace where that gap comes from.
We Started With the Assumption That It Was Prompting Skill
Whenever productivity gaps come up in AI collaboration, the most common prescription is prompt-writing technique — be more specific, assign a role, include examples. That prescription isn't entirely wrong. The same request phrased differently really does change the direction of an AI's output.
So our first assumption was that the difference came down to prompting skill.
The crack in that assumption showed up faster than expected. We watched what happened when people who'd thoroughly mastered prompting technique took on work in a field they didn't know well. Just changing the domain changed the results. The output had the right shape but felt like something was missing — structure without substance. Prompting skill alone didn't guarantee quality.
Another observation shook the assumption further. When people with proven prompting skills received AI output in a field they knew well, they revised it far faster and more precisely than average users — with the same prompting technique. That was a signal that the variable wasn't prompting skill at all, but something else.
We had to change direction.
What's Stopping People From Fixing AI Output
We went back to the people in the same experiment and asked what they did after they got their results. A common bottleneck emerged.
Some said, "Something feels off, but I can't tell what." Others said, "I know what's off, but I couldn't tell the AI how to fix it." Either way, their revision instructions turned vague — things like "make this sound more natural" or "make it more persuasive." The AI responded to those requests, but the output kept wobbling in the same place.
So we started collecting examples from the opposite direction — people who fix AI output accurately the moment they see it.
An editor who has edited more than 200 articles reads the third paragraph of an AI-written draft and immediately says, "The argument shifted register — it was making a strategic point and suddenly dropped into execution-level detail." That analysis goes straight into the next instruction. The AI responds to it precisely.
A director who has run marketing campaigns for nearly a decade instantly spots which of the AI's proposed directions repeats a pattern she's already watched fail. Because that judgment is fast, she settles on a direction within five minutes of seeing a draft. A copywriter finds the exact point where an AI-written sentence loses its rhythm and lays out three concrete ways to fix it.
What they all had in common was the ability to evaluate the output. Because they could evaluate it, they could give specific revision instructions — and because the instructions were specific, the AI returned useful revisions.
Watching this process revealed why everyone else was getting stuck. The bottleneck wasn't AI usage — it was evaluative ability. The difference was whether someone could say, concretely, what was good and what wasn't.
Where Does That Evaluative Ability Come From
This observation pointed in a clear direction. People who've built up judgment in their field can evaluate AI output and decide how to revise it. People without that judgment can't move forward even after the AI hands them something.
So where does that judgment come from? That question drove the rest of the investigation.
The editor's evaluative ability was built by editing hundreds of articles — repeated exposure to which paragraphs flow naturally into the next and where readers drop off. The copywriter's instinct built up the same way, from writing hundreds of pieces of copy and watching how people actually responded. The campaign director's judgment came from the same process — living through, firsthand, which campaigns failed and why, and which ones worked and why.
None of this came from studying theory. It formed through a cycle of making something, putting it out, watching the response, and revising. It's the same reason someone who's run a brand for years can immediately articulate their brand voice. The standard wasn't there from the start, fully formed — it emerged from staying consistent and building up pieces over time. Once that instinct exists, a person can judge new output on sight.
The conclusion this points to is an uncomfortable one. One reason AI collaboration doesn't work well may be that a person's own domain instincts haven't been put into words. Many people assume they can recognize good work when they see it. But the moment they have to hand that standard to an AI, the vague feeling often fails to translate into a concrete instruction — and it's easy to blame the AI, or chalk it up to weak prompting, when that happens.
In AI collaboration, no amount of prompting skill compensates for missing domain judgment. If you can't evaluate what's generated, you can't use it as a starting point to refine. The variable this investigation pointed to wasn't AI proficiency at all.
Figure Out Where You're Actually Stuck First
People struggling with AI collaboration often say the same thing: "I guess I need to get better at using AI."
Sometimes that's the right call — tool literacy really is the weak link in some cases. But if AI collaboration isn't working in a field where you already have real experience, it's worth checking a different possibility first: can you actually put into words what makes a result good in that field?
That check is uncomfortable for a reason. It's less about how to use AI and more about how much of your own domain instinct has actually been articulated. That work moves much slower than learning a tool. And without that instinct, no amount of prompting skill will stop the same problem from repeating — being unable to fix what the AI hands you.
When someone starts a new role, figuring out early what success looks like in that role shapes their first few weeks — you need to know the standard before you can act, before you even sit down at the desk. The same order applies to AI collaboration. The starting point is putting into words what "done" looks like for the task at hand.
This investigation doesn't lead to a complete answer. Understanding what AI tools are good at still matters, and prompting technique still has its uses. But if the feeling that "AI just isn't working for me" persists, it's worth checking whether the real bottleneck sits upstream of the prompt. Can you say why something the AI produced feels off? If that answer doesn't come, the place where learning needs to start may be somewhere else entirely.
