Somebody On Your Team Has The Good AI Prompts. That's The Problem.

Last updated: August 27, 2026

Somebody on your team has the good prompts. They found the workflow that actually saves four hours a week, they know which tool is real and which one is a demo, and none of it is written down anywhere.

A year ago that might not have mattered much. Now it does, because AI does not close skill gaps. It widens them, fast, and mostly where nobody is looking.

Molly Graham wrote the standing advice for this kind of thing years ago: give away your legos. Hand over the part of the job you’re best at, the part that makes you feel needed, so somebody else can build with it. I think that advice breaks a little when the thing you’re hoarding isn’t a role but a prompt library. Where it breaks is the interesting part, and I’ll come back to it.

Why does the AI skill gap open up faster than people expect?

Because velocity hides it. Everything looks fine right up until it’s just one person’s judgment holding the whole thing up.

A founder I coach put it better than I could. He looked at what a year of AI had done to his engineering team and said: “If they were just task processors one year ago, now they are just faster task processors.” He’d spent his own AI hours on domain research and understanding the business. His engineers had spent theirs on implementation. Velocity climbed every sprint. So did the gap between what he understood and what anyone else on the team understood, until he was the only person who could make a real call. Which is a lovely position right up until you want a weekend.

That’s the version that never announces itself. The dashboard looks fine. What’s quietly shrunk is how many people on the team can actually reason about the work, and it shows up a quarter later looking like a team empowerment problem you’ll spend weeks misdiagnosing, when the actual cause was one person’s head start compounding in private.

What does giving away your legos have to do with it?

Because Graham’s advice was never about handing off the part of the job you disliked. It was about giving away the part you were best at and loved, the part that made you feel like you were the reason things worked. That’s what made it hard. Nobody struggles to delegate the expense reports.

A prompt library is exactly that part, for a lot of people right now. You built it. You’re faster because of it. Handing it over is supposed to work the same way handing over a product area does.

It doesn’t, quite. I’ll get to why in a minute, once I’ve told you what actually closes the gap while it’s still small enough to close.

What actually closes the gap?

You reward the sharing, not just the thing it produced.

This is the part that actually fails in most companies. If the only thing that gets noticed is what somebody shipped, then handing over the workflow that let them ship it is a straight subtraction from their own standing. Nobody does that voluntarily. You have to say the quiet thing out loud: I want to hear about the thing you built, and I’m going to make a fuss about it in front of people, not just about what it produced.

I know what it costs to skip this, because I watched it happen somewhere I loved. I spent over a decade at Sonos, joining as its first software product manager and staying through the stretch from about 50 people to 1,500. We hired carefully for open minds and no ego, and we still ended up somewhere people were sometimes afraid to speak up. Looking back, that’s the thing I’d fix. It didn’t come from one bad decision anybody announced. It came from a hundred small ones about who got included in what, none of which felt like a decision at the time.

An AI skill gap opens the same way. Nobody decides to hoard a workflow. They just never get asked about it, the moment passes, and six months later they’re quietly running the department in their head.

What if you’re the one with the good prompts?

Then you’re the problem, and this is where Graham’s advice actually bites.

You built the workflow. You have the prompt library in a private note. You can get a usable answer in one pass while everyone else is on their fourth try. It feels like competence, and it is, and it’s also the exact shape of a bottleneck. Every hard call now routes through the person with the good prompts, and you’ve quietly become a dependency wearing the costume of a high performer.

Giving that away is worse than giving away a lego, because a lego actually leaves. Hand somebody a product area and it’s genuinely theirs now. Hand somebody your prompt library and you still have it, you’re still faster, and you have to keep choosing not to be the one who does it. You did not hand off the lego. You handed out lego, kept the box, and now have to keep your hands still while somebody builds it slower than you would.

That’s not one policy meeting. It’s a hundred small calls about who gets to move faster, made quickly, mostly by default, by somebody who didn’t think of them as calls at all.

Your tools can draft the strategy doc, summarize the research, and write the exec update. They can’t tell you whether the gap on your team is a training problem, or whether you’re the one keeping the box.

Frequently asked questions

What is the AI skill gap on a product team?
It’s the growing distance between the people who’ve found real workflows with AI tools and everyone who hasn’t, and it grows faster than most other skill gaps because velocity hides it. Output looks fine right up until one person’s judgment is quietly load-bearing for the whole team.

Why does AI make skill gaps worse instead of better?
Because the person who’s already ahead compounds their advantage every sprint, and nobody’s checking whether that advantage is spreading or concentrating. A founder I coach described his own engineering team a year in: “If they were just task processors one year ago, now they are just faster task processors.” He’d spent his hours on domain understanding, they’d spent theirs on implementation, and the gap between what he understood and what they understood kept widening.

What does “give away your legos” mean, and does it apply to AI?
It’s Molly Graham’s advice for scaling teams, first published in First Round Review: give away the part of the job you love and are best at, not the part you disliked, so somebody else can own it. It applies to an AI skill gap, with one difference that makes it harder. A lego actually leaves your hands. A prompt library doesn’t, so you have to keep choosing to let somebody else do it slower.

What do you do if you’re the bottleneck yourself?
Notice that being the fastest person with the tool isn’t the same as being the most useful person on the team. If every hard call routes through you because you have the workflow nobody else does, you’ve built a dependency, not a strength. The fix is the same one you’d ask of anyone else: give away the part you’re best at, on purpose, before someone has to ask.

How do you get people to share what they’ve learned about AI, instead of sitting on it?
Reward the sharing itself, not just what it produced. If recognition only ever attaches to the output, handing over the workflow that made it possible is a cost with no benefit to the person doing it. Say plainly that you want to hear about what somebody built, and make a point of it in front of the team, separate from whatever it shipped.

If this is happening on your team right now

If you can already say who has the good prompts, or you’re starting to suspect it’s you, that’s exactly the kind of call I work through with product leaders. Book a call and we’ll work through it.