ou know the feeling, even if you'd never say it in a standup. You spent ten years getting genuinely good at something, and now a model writes in thirty seconds what used to be your whole afternoon. Somewhere under the green dashboards there's a quieter thought: if the machine does the building, what exactly am I for?

I want to answer that, because the fear is real but it's aimed at the wrong target. The thing being automated is the typing. The thing that's appearing — and won't expire — is the directing. And the strange way I can prove it to you is by telling you how we just tried to hire for it.

A job you can't put in a PDF

We needed someone to run AI delivery — to hold several client projects at once by orchestrating agents and judging their output, not by hand-coding each one. We sat down to write the ad and hit a wall. Every word we reached for belonged to the old job. "Senior engineer." "Owns the codebase." It described typing. The role isn't typing.

So we did the on-brand thing. We didn't write a job post. We built one you talk to.

It's a public repo. You point your own assistant at it — ChatGPT, Claude, whatever you already use — and it reads the role, weighs your fit against what it knows about you, walks you through a real task, and tells you the truth, including a clean no. Not a brochure. A conversation, with an advisor who happens to already know you.

Why this works now, and didn't a year ago

It's been roughly a year since ChatGPT and Claude shipped memory, and Claude Code recently learned to remember a project across sessions. If you actually work with these tools, you almost certainly left memory on — because it makes every answer better. Which means your assistant isn't a blank intern. It knows your background, the things you've shipped, the way you reason about a hard call.

So when it reads our role and turns to you, it can do something no recruiter and no CV can: judge your fit against the real you, then coach you toward the parts you're light on. It's a better first-round screen than we could ever run from the outside — and it runs on your side of the table.

The privacy line matters, and we say it plainly: your context goes to your own AI provider, exactly like every other chat you have with it — and never to us, unless you choose to send the application it helps you write. We don't see your memory. We just see who shows up, and what they made.

“The people who recoil from this are exactly the people the role isn't for. They don't get screened out. They screen themselves out.”

That's the part I find quietly beautiful. A posting like this doesn't need a filter — it is one. Someone who keeps AI at arm's length, who turned memory off, who feels the dread and chose to look away, never gets far enough to apply. Someone who points their agent at it and leans in has already demonstrated the single trait the job is about. The medium is the test.

This is bigger than hiring

Step back from recruiting for a second, because the shape of this generalizes, and it's the part that should make a marketer sit up.

The next reader of your website may not be a person skimming your pages. It may be their AI, reading on their behalf, surfacing only what matters to them. Picture a travel agency with a hundred itineraries — the kind of catalog no human ever reads to the end. Now picture a visitor's assistant, which already knows the ages of their kids, the budget they actually have, the trip last summer that went sideways, quietly finding the one trip that fits and staging the booking. A brochure that reads itself, for an audience of one.

It's better for everyone in the loop. The visitor consumes a wall of information through the interface they already trust, enriched by a model that knows them in a way your funnel never will. And the personal context stays where it belongs — with them and their provider, never dumped into your CRM. You don't have to surveil someone to serve them well anymore. You publish for their agent, and let their agent do the personalizing you used to fake with a cookie banner.

We started with hiring because that's where we needed it. But "publish something a person's AI can read on their behalf" is a primitive, and most of marketing is downstream of it.

The moat is the thing you were afraid for

Now back to the dread, because here's where it inverts.

The reason this role doesn't expire is the same reason it's hard to write an ad for. It runs on the parts of you that don't compress into a weights file — the judgment that doesn't parallelize, the taste you can't fake, the experience that tells you which of five plausible plans is the one that holds. A better model doesn't replace the orchestrator. It walks into the house you designed and does the work better — which only makes you more valuable, not less.

Software development as we knew it is changing, yes. But you are still the person your colleagues bring the hard decision to. You're still the one in the room when the call gets made. That seat doesn't get automated, because the machines are what's being directed from it. That's the moat. The typing was never it.

If you've felt the fear, this is the doorway out of it — not "learn to prompt harder," but convert what you already are, a builder with judgment, into the one role that gets stronger every time the models climb.

Read it with your own AI

The whole thing is open and MIT-licensed — github.com/orangehill/orange-hill-orchestrator-agent-mode. Point your agent at it. There's a file for the role, one for the task, and a short stack of company files so it can represent us honestly.

The one to read closely is company/orange-hill-dna.md — distilled from these very essays, it's what we stand for and how we actually operate. It's the lens we want your agent to use when it tells you, for real, whether this is you. And if you run a company and you've felt the same wall we did — that the old words don't fit the new work — fork the whole thing. That's the point of leaving it open.

If it is you, the role is here: orangehill.ai/en/careers/orchestrator. Bring your own AI. We reply to everyone who does the task.