Stencil the Values on the Wall
Who has the longest table of them all.
The week of Config, I was in San Francisco, spending more time in the biergartens that gather around the conference than in the conference itself. I met so many kickass people, doing incredible and brilliant things. In one of them a designer told me she had learned something new every month for a year. New tools, new workflows, new ways of working next to a machine that could suddenly do a version of her job before lunch. And she was good at it! But she was also tired in a way that had nothing to do with the hours. What she wanted was not another tool. She wanted someone to tell her what good looked like now, so she could stop guessing every single time.
She was not the only one. Across those rooms I kept meeting the same person in different clothes. People building small advocate teams inside their companies, trying to earn back enough trust that anyone would contribute a pattern to the shared system again. People at large organizations working out how to use these new tools without losing the soul of the design, while the budget conversation got harder in the next meeting. Passionate, careful people, most of them out of time, all of them lowkey asking the same question. What does success even look like when the work happens this fast, and how do I know when I have it.
On the actual stage, two of the most‑talked‑about sessions looked like they were arguing about exactly that.
Meta’s Harvey Whiting made the case for building a system a machine can read. Specify everything. Write the golden sets. Treat ambiguity as a defect, because a model will fill any gap you leave. “Anything that is not specified,” he said, “is randomized.” The counsel underneath it was simple. If you want speed and quality at the same time, the quality has to be written down somewhere the machine can find it, or the speed just gives you more of the wrong thing faster. Speed without quality, Whiting said, is expensive noise.
A few sessions away, Mercury’s team argued something that sounded like the opposite. Lauren LoPrete called her talk design systems anarchy. Stop policing. The gatekeeping era is over. She described the new tools as their Xerox machine, a copier that let anyone in the company reproduce the house style without asking permission. Her design system had stopped being a rulebook and started being a culture. Lower the floor. Trust people. Get out of their way.
Both of them were right, and both were describing something they had watched go wrong. Meta had seen speed without structure turn into a firehose of plausible, slightly incorrect work that somebody then had to catch. Mercury had watched a design system become, in her words, a bouncer at the door: a checkpoint that slowed every project down and trained good designers to ask permission instead of make decisions. Neither of them was wrong about the failure they were fixing. They had just fixed it from opposite ends and come out sounding like rivals.
They were arguing for the same thing. You could hear it most clearly in Mercury’s own best example. LoPrete talked about 924 Gilman, the punk club, where the house rules are painted on the wall.
No racism. No sexism. No violence.
Everyone who walks in can read them, and everyone agreed to them by walking in, so the room enforces itself. She told it as a story about culture beating control. The culture runs itself, she said.
Sit with the mechanism for a second, though. A short list of values, written down in a place everyone can see, enforced by the room instead of a person at the door, is a design system. It is the most legible, most machine‑readable thing in the building. The wall is doing the governance. Nobody has to stand next to it.
Structure is how a culture runs without anyone standing guard. When the wall is legible, you do not need a bouncer, because the wall does the enforcing. When it is not, someone has to stand there and do it by hand. Usually that someone is the person who cares the most.
The designer in that room in San Francisco was tired because she had become the wall. Every drifted component, every off‑brand exception, every “can you just take a quick look” came to her, because the values of the system lived in her judgment and nowhere a new hire or a machine could read them. She was not failing at her job. She was doing the job of a system that could not do it for her. A value nobody can read is a value someone has to enforce by hand, and she was the hand.
This is where the new tools stop being an abstract worry and start mattering in a specific way. Hand an agent a design system it cannot read, and it will do exactly what everyone at Config was nervous about. It will produce a great deal of confident, fluent, slightly wrong work, at a speed no human reviewer can keep pace with, and every piece of it lands back on the person who holds the real standard in their head. Hand that same agent a system where the line is actually written down, where the values are specified in something other than one exhausted person’s taste, and it will hold the line for you at three in the morning when no one is watching. Same tool. The only variable is whether the wall exists.
So the work is to make the values legible.
To people, so a new designer knows the line without being told twice and can contribute without fear of tripping an invisible rule. To the machine, so it can carry the standard instead of manufacturing more work that violates it. You hold one line hard and you leave the rest loud and open, the way Gilman holds a short list and lets everything else be punk. You lower the floor so more people can build, without lowering the ceiling on what good means. And you let the artifacts make the argument, so the system teaches the standard instead of guarding it.
There is a harder edge to this, the one that surfaced in every budget conversation I sat near in those rooms. When the economics shift (and they are shifting) the people who hold the standard in their heads become the most expensive and the least legible part of the system. If the values live only in them, then losing them, for any reason, means losing the standard along with them. Writing it all down where a person and a machine can both read it does more than make you faster.
It is how the soul of the work survives a quarter when the headcount does not, and how a team earns back enough trust that people contribute to the system again, because a wall you can read is a wall you can build on without fear of what you cannot see.
None of that replaces the designer. It gives her judgment somewhere to live besides her own memory and her own evenings. That is the whole reason I started HeyHi. Not to automate the person in that room. To build the infrastructure that holds the standard with her, so she stops being the only thing holding it up.
Extend what a team can do. Never replace how they think.
Config has come and gone, and the argument from those two talks has not settled. Which is why I would rather put this out while it is still live than six months from now as a recap.
The people I met in San Francisco did not need another take on whether structure or culture wins. They needed a place to start that was not the blank page. Most of them were sharp and generous and completely out of hours, because nothing yet existed that they could pick up and build on. Building the wall from nothing, on top of the actual job, is how you end up too tired to paint anything on it.
So we built ours in the open, and we have been handing the pieces out as they were ready. One is already out: agent-corrections, a small thing that does one job. It lets a mistake, once named, stay named, so an agent stops repeating it and a person stops catching it by hand. One value, written on the wall, holding itself.
Today the whole scaffold goes up. I called it Keel: the way we structure context, skills, and memory so a single person can operate like a team without losing the soul of the work. That corrections registry is one piece of it. Both are free, MIT licensed, no strings, nothing gated behind a call. Keel is not the finished answer to anything. It is a wall you can start from, so you are not standing guard over an empty room, hoping you remember every rule yourself.
Because it is unfinished, it is open. Keel ships with a handful of examples to learn from, and is otherwise yours to fill, waiting for your patterns and your corrections, not ours. It is built to be added to, and how to do it is written down inside. If you work out something that holds a line we missed, send it back, so the next person who walks in tired and out of time inherits your work instead of starting from nothing. That is the version of this I actually want: one where the table is built longer, rather than a wall that continues to rise.
I keep thinking about the designer who just wanted to know what good looks like now. I do not think anyone can hand her that answer. I think we can hand her a starting point, a little less fear about the next step, and the room to keep pushing at the edge of what this work can become. The wall does not have to be finished. It has to be legible, and it has to be there. You can build on it from here.