The Depth Penalty Is Gone
I spent five days inside a volleyball club's operations — WordPress, a marketing CRM, a registration API, consent law. None of it is my domain. That stopped mattering, and the reason it stopped is the interesting part.
Nino Chavez
Product Architect at commerce.com
Forty-nine of the ninety-five programs in a volleyball club’s registration system were past their end date. The system reported every one of them as live. Twenty-one were still accepting sign-ups — a parent could start registering their kid for a session that had already happened.
I found that in the first week of an engagement I had no conventional business taking.
Here’s the point, up front: the thing that used to stop me from taking it — the depth penalty, where a generalist can either be broadly useful or go deep in an unfamiliar domain, never both on a deadline — is gone. Experience steers. Agents execute. What follows is what that looked like on one real engagement, including the two places the machine was flat wrong and the number I still don’t have.
None of This Is My Domain
My day job is enterprise commerce architecture. The club — 630 Volleyball, whose photography and media I already handle — needed something else entirely: a website that had gone fifteen months without an owner, a marketing platform whose outbound half had never been turned on, a registration API, and marketing-to-minors consent law. None of it overlaps my day job.
Five days and 222 commits later, there was a rebuilt thirty-eight-page website verified on a private copy, a read-only sync layer pulling live program data from the registration system, an operating model, a consent policy that refuses to publish when it isn’t sure, and an evidence register — a document that separates what we verified ourselves, what someone told us, and what we’re still guessing.
The typing wasn’t mine. Agents wrote the plugin, ran the audits, read the API docs, drafted the standards. The judgment was mine, and it turns out judgment was the scarce input all along.
Steering Is Knowing What Would Be Embarrassing to Be Wrong About
Steering gets described as prompting. It isn’t. On this engagement it was a short list of standing rules, applied every day:
- Every number is sourced inline or labeled invented. No third category.
- “Verified” means we pulled the source this session. Reading an old claim of verification is not verification.
- Nothing a family sees ships without a human tap. The consent rule fails closed — when in doubt, it refuses.
- A count is never the result. Registrations and reachable families are the only numbers allowed in the headline; everything else is a diagnostic.
None of those rules require knowing WordPress. They require knowing where projects like this die: an unsourced figure in a client document, a “verified” that nobody verified, an automated post naming a twelve-year-old without permission.
Steering is knowing what would be embarrassing to be wrong about.
That’s what twenty years buys now. Not the typing — the flinch.
The Problems Nobody Asked About
The ask was modest: pages that stop going stale. The finds that mattered were things nobody had the vocabulary to ask for.
The sharpest one: the off-the-shelf integration template between the registration system and the marketing platform — the thing any reasonable person would click to connect the two — would have quietly created marketing contact records for families of minors carrying no consent signal at all. The opt-in lives on a different record than the one the template reads. One default away from marketing to families who never said yes, and no error anywhere.
The club also had 163 conversations where a family wrote last and nobody answered. Six phone calls in six weeks, none picked up. Hundreds of website form submissions filing into a database nothing reads.
Nobody hid any of this. It was invisible because seeing it requires reading systems nobody was staffed to read. The agents read them in an afternoon; experience decided which findings were load-bearing.
Where the Machine Was Wrong
Two failures are worth the space, because both were caught by discipline rather than expertise.
Every date on the rebuilt site rendered one day early. A timezone bug — and the only reason it surfaced is that seventy of the programs carry their date in their own names, so the page contradicted itself. If the data had been slightly less redundant, we’d have shipped it.
Earlier, a check reported that the registration API carried “no monetary amount anywhere.” False. It had searched for a field named like a price. The re-check enumerated every field and found fees on 84 of 95 programs — and also found the real answer: several amounts can apply to one program, and no field anywhere states what a family actually pays. So the site links to registration instead of restating prices. The wrong claim and the right design decision came from the same question, asked twice with different discipline.
And the honest scope limit on everything above: none of it is sold. The club has proposals, not invoices. Producing in five days what used to take a quarter proves production, not value — which is why the engagement gates the sale on a one-week measured trial before any monthly fee exists.
The Two Numbers I Don’t Have
The trial measures two things: how many minutes a day the club’s director actually spends approving, and how many drafts go out without edits. Good numbers become the contract’s standard. Bad numbers kill the pitch — that’s not a risk in the design, it is the design.
Until then, the claim stays exactly this size: an outsider steered agents into an unfamiliar domain and came back in five days with working software, receipts, and problems the client couldn’t see. The fix for those forty-nine phantom programs exists on a private copy of the site, reading the real dates, waiting on a yes.
The depth was never the hard part. It just used to be the expensive part.