Notes
Sorry, Raymond.
It is time to take the shell off.
In 1929 a manufacturer asked Raymond Loewy to make his duplicating machine less ugly. In three days Loewy designed a shell. It covered those machines for forty years, and it more or less founded industrial design.
That was the right call. Machines then were, in his words, enormous and sinister looking, and the mechanism was no use to anyone. Design covered it up and handed people something they were not frightened of. We have been doing it ever since. And it was fine back then, as a duplicator did the same thing on Tuesday as on Monday. Hiding it cost nobody anything.
And crucially, somebody understood how the engine worked, by opening the shell. You could take the cover off a duplicator, follow the linkages, and account for every part. The shell was a courtesy, not a secret.
Designing the product demands taking the shell off, and designing how the model behaves.
Because what sits under there is no longer knowable by inspection. It is assembled out of something you did not train, on data you have not seen, and it shifts as it goes. The shell still goes on, in the same reassuring way. But increasingly nobody in the room can take it off and account for what is underneath — not the user, and not the team that shipped it.
And this is what I did in Furlong
Furlong is an app I built. It reads your running history from Apple Health and each morning tells you one thing: what to run today, or to rest, so that you keep running rather than getting hurt.
As part of the design process I built a few interactive visualizations that expose how the system works, and how its predictions are made in an environment that keeps changing based on a runner’s workout load. The one below is one of the ways I lifted the shell and looked at how the engine actually behaves.
Signals on the left. The rules in the middle, checked in order. The week that comes out on the right. Click any day to see which rule produced it.
Families of rule
- Health. Outranks everything. No distance is offered against it.
- Recovery. A race, a layoff, one unusually big day behind you.
- Rhythm and load. Days in a row, days this week, the ratio.
- The key session. The long run, and the rest day that protects it.
- The plan working. Nothing to flag; build the base.
Why the order is the model
The ordering is the safety argument, not an implementation detail. A sore shin outranks the schedule; the schedule outranks the load numbers. Anything above a rule can override it.
So there is no score, and no way for a good ratio to argue its way past a symptom. It also means two identical-looking weeks can come from the same rules reaching different answers, because a different rule got there first.
What it never does
It does not diagnose. The niggle rules route you — bone, tendon, muscle behave in opposite directions and are managed in opposite ways — and then get out of the way.
It does not chase a missed session. A long run that does not happen comes round again; it is never squeezed into the following day.
What lifting the shell found
Drawing it is how I found that the system needed improvements: missing long runs, and several other things the code looked perfectly reasonable about.
Seeing a whole week at once showed that a handful of rules were deciding almost everything, and that several never fired at all. Running it forward over months showed worse: a plan that quietly stopped scheduling the long run, and a safety cap for beginners that shrank the very people it was meant to protect.
You cannot read behaviour. You can only watch it. It took eleven versions of a similar picture, run across sixteen weeks, before I stopped patching and rewrote the engine — that one is here.
Then the tool became the product
The app itself now shows its work. Every suggestion carries its reason, and you can open it to see what decided the day.
What I built to see the machine turned out to be the product.
Loewy’s shell was a generous thing. It took a machine people found frightening and made it approachable, and it cost nobody anything, because whoever needed to could still lift it and see the linkages underneath.
That last part is what we have quietly dropped. The covers kept getting better. The ability to lift them went away.
So the brief has changed. It is not to put a nicer shell over something nobody can open. It is to keep the thing openable — by the people building it first, and then by the person being asked to trust it.
The shell was always half the job. The engine is the other half.