Everything to everyone, now hiring

I keep telling it the brand isn’t green.

For about a week now I’ve been building a design system in Claude Design. The palette has no green in it. Not a shade, not a hint. And the tool keeps reaching for green.

So I tell it. It agrees right away, fast and a little too gracious. “You’re right, that’s a real inconsistency, good catch.” It swaps the green for blue. Then it builds the next set of components, and the green is back. I tell it again. No green, it’s blue. It agrees just as readily, fixes it again, and the green returns in whatever it builds next.

Five rounds of this. Probably more. I stopped counting.

The story everyone is telling right now

I keep running into a version of the same argument. The agency, or the team, or the one person working alone, becomes a kind of overseer. The AI does the work. You supervise it. One human overseeing a row of agents, the way one attendant covers six self-checkout lanes, stepping in only when one of them starts flashing for help. Work pours out one end of the line and you stand at the other end, catching the exceptions.

It’s a clean story. It assumes the system holds still.

The system does not hold still.

The fix never sticks

The green looks like a capability problem. Partly it is. But the part that actually matters is that the fix doesn’t stick and doesn’t spread. It corrects the exact thing I pointed at, then makes the same mistake two components over, in work I haven’t looked at yet. Catching the error doesn’t end anything. There’s no point where I’m done checking. Nothing I corrected yesterday is still corrected today.

And green is the easy version. Green is visible. I can see green.

It loses the thread

The structure drifts the same way. I’ll set up a design system, and a few days later the tool has quietly stopped matching it. So I stop and tell it to go read its own system again, look at what already exists, realign the new pieces to the old ones. I’m doing exactly that this week, on work I started last week. That isn’t a step in the process so much as a second one running underneath the first, all the time, just to keep the thing from coming apart.

Then it changes your voice

The one that actually worries me doesn’t happen in a design tool. It happens in writing.

I’ll work on copy for a business and the voice will be right. I come back to it later and it’s moved a little. Not wrong. Just not the same hand. Keep working and it keeps moving, a few degrees at a time, every step reasonable on its own, until the thing I wrote last week and the thing I wrote today don’t sound like they came from the same company.

You don’t catch that in the moment. There’s no green to point at. The drift is slow and quiet and every individual step looks fine, and the only way to see it at all is to hold the whole arc in your head and feel the distance between where you started and where you are. Which means the only thing standing between a business and a slow slide into somebody else’s voice is a person who remembers what it was supposed to sound like and is paying close enough attention to notice it move.

I only caught any of it because I knew where to look

Every one of those is a lane I actually know. I can fight about the color because I know the palette cold. I can feel the voice slide because the voice is my own work. Take away that expertise and I lose the fight before it even starts.

Because the same story puts one person over the entire line. Web, SEO, PPC, social, outbound. Take PPC. I’ve run paid campaigns. They failed. Nearly every ad I’ve put up has underperformed. Could I get good at it if I sat with it long enough? Probably. But I haven’t, and I’m not going to, because I have a business to run and other work that actually needs me. So I am not the person who can look at a campaign the tool just built and know why it’s quietly bleeding money, or catch that it pushed something live that should never have gone out.

That’s the honest version of the overseer. Someone who has touched a few of these disciplines, failed at one or two, and has no time to master the rest. So when the tool hands me a keyword strategy and tells me, with exactly the confidence it used on the green, that it’s solid, what am I checking it against? In the lanes where I don’t have the eye, the same drift and the same sure-footed mistakes are happening, and I’d have no way of knowing. The output looks finished. It always looks finished. Looking finished is the one thing these tools are reliably excellent at.

When you can do everything, you do nothing especially well

The part that gets me runs straight into the one thing this industry actually agrees on.

For years the advice has been to niche. Specialize. Get known for one thing and get good at it, because the generalist loses to the specialist every time. You get known for one of them, branding or paid or social, and you get good at that. Not all of them at once. The person who claims all of them is usually the person you don’t quite trust with any of them.

The overseer model quietly reverses that. It asks one human to be fluent in web and SEO and PPC and social and outbound at the same time, good enough in every lane to judge whether the machine’s work is any good. That is the most extreme generalist anyone has ever drawn up, expected to be sharp enough in all of it to grade the machine in all of it. It’s a job description that amounts to being everything to everyone.

I spend a lot of my actual work talking people out of being everything to everyone. It’s the most reliable way I know to become forgettable, and to do a pile of things at a level that’s fine while doing nothing at a level that wins. Now the same instinct is being handed back to me as the future. One person, every discipline, all at once. When you can do everything, there’s almost always nothing you do really well.

Two jobs, five lanes, no end date

So when people call the human here the overseer who steps in to catch edge cases, the framing is off in more than one direction.

These failures don’t sit at the margins. They’re constant, and they run through the middle of everything. That’s the first problem. The second is that catching them assumes I can see them, and seeing them assumes I know every lane well enough to tell real work from work that only looks real. The third is that even if one person could somehow hold all of it, holding all of it a little is a long way from holding any of it well enough to catch the quiet, expensive mistakes.

The story is asking one generalist to be two things at once. The eye that can judge the output across every discipline, and the memory that holds each one steady over time. Across five lanes. With no end date. Neither of those is a small thing you do between the bigger things. On work that has to stay consistent over time and hold up across disciplines, the eye and the memory might be the entire job.

More capable isn’t the thing they’re betting on

Maybe the tools get better at holding a line. They probably will. But capability isn’t the axis this model actually leans on. It’s leaning on consistency over time, which is a different thing. And it’s leaning on one human being able to judge every lane, which was never the model’s problem to fix in the first place. A thousand times more capable still doesn’t make me a PPC analyst, and it doesn’t hand me the years it would take to become one across five disciplines at the same time.

A more fluent tool doesn’t drift less. It drifts more convincingly. It tells me it fixed the green with even more confidence, and then it serves me green again.

I’ll let you know when it stops. I’ve been telling it for a week.