Skip to content

Updated 10-19-2026 · v0.3.0 · Mountain View

· Essay

Whose skills are they

Your taste is now a file. Files have owners. Taste does not.

Cut-paper collage in bright colors: a bird in orange, blue and yellow among leaves, flowers and shapes on a white ground.

"Your taste is now a file. Files have owners. Taste does not."

Encoding the wisdom of hands and minds

Skill used to live in the hands: a potter knowing when to pinch the clay and when to let the wall rise, a cook knowing how much salt sits between three fingers, a baker knowing whether a loaf is done by knocking on it. Nobody could write that down, including the person who had it. You had to study it by standing next to someone for years.

Knowledge work moved the skill up from the hands to the mind and gave it new names: hard skills, soft skills, judgment, taste. The essence was the same - knowing how to do something, and knowing when it is done - and it still went home with the person every night.

In digital design specifically, skill meant knowing customers deeply, how the business makes its money, and the craft itself: which dial to reach for, how to critique an artifact, when a composition is done and how all of it serves customers' and business needs. Over time, the expertise compounded for both the company and you, and the only way for a company to keep it, was to keep you and encode what it could into process and product.

Today, the encoding goes a step further - to get an agent to work the way you do, you codify it: a skill file, a memory store, a harness config, an eval rubric, the reference you train and score against. Tacit becomes explicit. Explicit is a file, and a file has an owner who is no longer you.

Who owns codified wisdom?

A file made at work is work product, per most employment agreements. It now covers the context and skills you have crafted over years, as you turn your biological neural network into an artificial one.

The overlap is where it hurts. "How I critique a design" is your craft plus the company's IP and processes, and it lands in one memory store: the customer data and the business logic came from the company, the insight and craft came from you, and the skill is all fused, and very hard to separate.

Giving your Legos away. @molly_g went back on @lennysan's podcast in September 2026 to amend her own "give away your Legos" advice in the age of AI: "keep judgment, vision and taste for yourself (instead of giving it to AI)" (paraphrase; Graham). I'd argue you owe the company your vision, judgment and taste. But how you arrive at the vision, how you think about step changes in a product or an ecosystem, is the Lego you keep.

What is needed is distillation at the moment you write it down, and keeping the method and conditions clean of the specifics.

The third owner

There is a third owner in the room. The agent that knows how you work runs on a lab's model. And in most cases it does not belong to the employer nor to you. @rsarver made the personal version of the argument in September: "the more a personal AI knows about you the better it works, which is the reason not to let a platform with its own business model hold it" (Sarver). At work the same memory has two landlords, the employer and the lab underneath it. You are neither.

That makes it a three-way standoff: you, the workplace, and the model. Each wants something different from the same file.

  • The employee - wants it to capture their taste so their talent compounds.
  • The workplace - wants it to work 24/7 and stay when the employee goes.
  • The lab - wants it to run best on its model, so that leaving costs the other two, and it wants to train on the file, because every skill written for its model makes the next model smarter.

A standoff holds because each side has something the others need.

  • The employee - holds the judgment that keeps the file current, and the parts that never made it into the file.
  • The workplace - holds the customers, their data and the business knowledge.
  • The lab - holds the model and runtime. A skill file does nothing without a model to read it.

Nobody can walk away.

Three overlapping circles titled “Most skills are written where all three overlap”: you, the employee (your own model); the workplace (company context); and the model (the model and runtime). The pairwise overlaps are the company's model, a personal account, and after you leave; the center is the work agent, where most skills are written.
Three parties, one file. Each overlap is two of them siding against the third. The centre, where all three overlap, is the work agent.

Firms are already paying to leave out the third party in the standoff: you and the workplace, with a model of your own. Ben Affleck's InterPositive would not build on labs' video models as they are trained on other filmmakers' work. Instead, for each film they train a private model on its own dailies (Hollywood Reporter). In Affleck's words, "they benefit from the learning that they're creating" (transcript).

Netflix bought the company in March 2026 for $587 million in cash (Netflix 10-Q, via Pondero) - That is the price of the studio-filmmaker alliance, and a hint at who owns the IP.

The standoff can be settled, but always with a price. Either you pay the hefty sum for a 1P model, or you use a 3P model and pay with your IP.

Encoding a bird's song

Whoever owns the file, holds just one point in time. The best line in this week's debate on codifying taste came from @mikeswartz_: "they're trying to cut open the bird to find the song".

Taste is a choice drawn from everything you have lived and weighed. A skill file holds one reading of it, from one moment, in one context.

A model generalizing this skill, averages it toward the flat middle whereas taste lives on the jagged edges - where inputs are sparse and nothing has been validated yet. Which is exactly where a prediction machine fails.

Back to the legos - Graham's worry was that handing the Legos to an AI agent stops the growth. But truly it's only a moment in time. One part of the loop.

Thirty years ago Ikujiro Nonaka watched Japanese firms turn craft into product and called it the SECI spiral: tacit to explicit, and explicit back to tacit. The four steps have names: Socialization, learning side by side; Externalization, writing it down; Combination, merging it with what the firm already holds; Internalization, absorbing it back and starting the next turn.

His example was a Matsushita engineer who apprenticed with a hotel's head baker to learn how to knead, then wrote the technique into a bread machine (Nonaka, HBR).

Writing it down was just one turn of the loop, not the end of the baker's skill. And with an agent, you grow by watching your own method run without you: every session shows you where it is vague, where it is wrong, and where it was never really written down. The spiral still turns. What matters is the evolution, and the evolution is driven by the human who can see the edges.

Where to stand

Build your skills IP. Designers always carried a portfolio out the door. Now everyone does, and it is a folder of skill files.

Four moves:

  1. Separate the containers yourself, before the policy does. Personal craft in a store you own, company process in the work agent. The test at write time is the Lego rule from section two: the outcome belongs to the company, the method is yours, so keep the method clean of the business IP and specifics.
  2. Keep the eval rubric portable. The rubric is the skill; the runs are the company's.
  3. Keep the lab replaceable. Write down what good looks like, not how to coax this model, and once a quarter check that a second model can pass your rubric. This one is not easy but important as some models require different architectures.
  4. Publish. Share the Legos: whatever you would be sad to lose when you leave, stand a clean version of it this quarter.

The firm keeps the last version. You carry the next.

Sources