Numbers travel. Arguments don’t.
In Taste is triage I said you can’t read every line anymore, so you feel your way and you judge the output, and whether it comes out as craft or slop comes down to whether you can say why.
This is the part I couldn’t answer.
There’s a hole in “judge the output.” How do you judge an output when you can’t see the argument that produced it?
I’ve been chewing on that for a while. I think I know the answer now, and I think it’s bigger than AI. But I didn’t get it the first time it happened to me, and I didn’t get it the second time either. If I hand you the conclusion now, I’ve done the exact thing this is about. So: three times it happened to me, and each one cost a little more.
The playbook nobody ran
The first one’s mine, and it’s the most embarrassing.
Before Genie, there was a document. We knew how to run discovery. We’d got it wrong so many times that we finally got it right, and we wrote it down properly. It was good. I wrote a lot of those pages, and I was proud of it.
Nobody ran it.
Not laziness. A document is something you have to remember to go and get, and keeping it current was nobody’s job, so within a few months it described a process we no longer had.
The playbook wasn’t wrong. It was unreachable. All that reasoning, sitting in a doc nobody opened at the moment they needed it.
I wrote the reasoning down somewhere nobody could reach it.
The why behind the tools
The second one’s worse, because I did it to somebody.
Genie gets built. It works. It does the thing I wanted it to do: take context and turn it into a prototype. I’m pleased with myself, and I hand it to the team.
A colleague takes it and doesn’t use it the way I built it. He pulls it apart and rebuilds it around how he actually works. My first instinct, honestly, was that he’d missed the point.
He hadn’t. He’d found it. He told me it was too complex, and he was right. At the time it was a pile of slash commands. So I made it simpler, and that’s when other people started using it.
But that’s not the line that stuck with me. He said he wanted the why behind the tools, rather than just being handed prompts to consume.
I’d handed him two years of decisions as an output, with none of the argument attached. I had no right to be surprised that he didn’t trust it.
I handed over the conclusion and kept the why.
The one that cost something
The third one’s the expensive one.
It’s the audit from Taste is triage: the deals site, and the deal cards that went from three on screen to one and a half. Every number in that audit was correct, and the recommendation was backwards.
I want to be precise about the failure, because it isn’t the one people assume. The machine didn’t hallucinate. It didn’t invent anything. That constraint was sitting right there, in the discovery notes and the interview transcripts. We heard it. I knew their business.
The argument was three steps back. The reasoning was in the system. It just didn’t survive being compressed into a score.
I have never been the product designer who wields data like a sword. I don’t have that mind. AI has allowed me to become that, and for that reason I put a lot of confidence, and hubris, and trust in numbers.
One shape
Three stories, one shape. The why was unreachable, it was withheld, and it was compressed out.
Every handoff drops a little of the reasoning behind the thing being handed over. Sales to delivery. Discovery to design. Design to engineering. Engineering to QA. Us to the client. Nobody lies at any step. I’m not lying, you’re not lying. The reasoning just doesn’t survive the trip. By the time it reaches you, you’re holding a number, and the argument that produced it is three steps back.
And here’s the uncomfortable part: none of those three leaks were caused by AI. The playbook was a document. The second one was me, talking to a person. The audit had all its evidence in a transcript a human being sat through.
AI didn’t create the leak. It industrialised it. The handoffs got faster, there are more of them, and each one still drops a little.
And now everybody can produce. I watched a delivery manager build a prototype that captured every requirement. Genuinely useful. And it looked like crap. That’s not on them; they did the right thing. The reasoning about what good looks like was just nowhere they could reach. On most projects there won’t be a designer staffed full-time. So the honest question for my discipline is: how are we in the room when we’re not in the room?
Your version of that question is different. What’s the one number you hand somebody that has lost its argument? Sales, maybe it’s the estimate. PMs, the ticket. Design, the spec. QA, the pass rate. Each one is a conclusion travelling without its reasoning, and whoever receives it can’t tell a good one from a bad one.
The slot is empty
What makes me think this is fixable, and not just true, is that we already built the place to put the reasoning. It’s just empty.
When we hand a component into a build, the handoff is deliberately small: the spec, the markup, the styles. That’s what travels. And in the spec there’s a field for exactly this: what the component must do for accessibility. The why, not just the what. There are fields for the constraints the implementation has to respect, and for what the DOM has to look like.
When I counted, there were a hundred and twenty-nine component specs. The accessibility field was filled on six. The constraints on seven. The DOM on four. The contract exists, it’s answered almost nowhere, and nobody noticed, because nothing checks it.
I felt that hole on a proposal. A colleague handed me a prototype that captured every requirement, with annotations and integration notes and the things the client had said in meetings, relayed back to them. There was branding, but it wasn’t the branding. Most of my time went on the hero. If the skill that built it had seen the design spec, I’d have spent that time with him instead, working on how we were going to wow the client.
Same hole, two ends. We answered the questions, wrote real contracts, and left the answers in a document next to the spec instead of inside it. A contract you answer but don’t emit is documentation. It isn’t delivery.
If the reasoning isn’t in the artifact that travels, it doesn’t travel. It doesn’t matter how good it was. The document is where reasoning goes to be right and useless.
The spec is the context
More of what we ship is now made by something that reads a spec and nothing else. It won’t sit in the discovery call. It won’t overhear the client say the thing that changes everything, that space doesn’t matter and density does. It gets the artifact, and that’s all it gets.
So the spec isn’t paperwork anymore. The spec is the context.
We were always losing this. We just had people in the gaps, quietly patching it from memory. At this volume that doesn’t scale, and the person doing the patching is the one who can least afford the time.
The difference is handing someone a picture of a button versus handing them the definition of one. Every rule a machine can check becomes a check, so people spend their taste on the calls no check can make. We built tokens so nobody has to re-decide what blue is. This is the same move, one level up.
I want to be in the room even when I’m not in the room. Not so I can block things. So that when a developer spins something up for a ticket at 11pm, or someone needs a prototype for a pitch tomorrow, what they get is already close.
Say why
Three things, all small.
I look for my empty field. Every discipline has one: the place the reasoning was supposed to go, where there was no slot, or there was a slot and we left it blank. Mine was both ends.
When something arrives already quantified, I treat that as the moment to ask why, not the moment to relax. It feels good to get a number. That’s my scar: a number is what a lost argument looks like.
I put the reasoning in the artifact, not beside it. The spec, the ticket, the PR description. Beside it is the same as nowhere.
What I didn’t understand when I first said taste is triage is that saying why isn’t something you do at the end. You can’t rebuild an argument that got dropped four handoffs ago. By then you’re holding a number, and you’re confident in it, and the argument is three steps back.
Numbers travel. Arguments don’t. The number reaches you. The reason it’s right doesn’t. So you have to put it in on the way through, every time, in the thing that moves.
We’re all about to be handed a lot of things to like. Say why, and write it down where it travels.