Anthropic published a report this month on how employment is changing inside the company itself, How employment is changing at Anthropic. One figure stayed with me: in an internal survey, 20% of employees said their job has shifted from doing the work to managing Claudes. Not managing people. Managing Claude sessions, the agents you run in Claude Code, Cowork or a chat, and judging what comes back.
When the company that builds the model spends a growing part of its day reviewing output instead of producing it, that says something about where the rest of us are going.
Building got cheap first
Producing used to be the hard part. Most of the cost of a feature, a screen or a text sat in making the thing. That cost has collapsed. AI writes the first version of the code, the copy, the research summary, the test plan. In product design, the question is no longer whether we can produce something. It is whether what we produce is worth shipping. Felix Rieseberg makes a related point in How to be more ambitious: the models can already do more than most of us ask them for.
For a long time at Qonto, finding an email we had sent to our users was not a simple task. The backend gave you a list of code names and no way to see the content: you could not tell what an email was for, or what it said. So I built Message Explorer. The first version took days and covered the emails under Content Design's craft, the backend ones; the CRM emails came later, when we wanted a wider view of everything we send to users. Today it shows you the localization key behind each line and links you straight to it, so you can fix the copy right there. It was made for Content Design, and CRM, marketing, growth and engineering now use it too: last month 111 people used it, previewed 1,266 messages and ran about 300 searches, and someone uses it every day. Search time has dropped for everything: new features, errors, typos, legal and risk and compliance changes. The next step grows out of the same idea: following the flow of sends through a customer's journey. The build was never the hard part. Days, for a working tool. Knowing what it had to solve was the work.
Judgment and taste become the edge
Judgment is knowing what good looks like before you see it, and being able to say why. Taste is the same instinct working earlier, the standard you hold before anything exists. Together they do the work production used to do: deciding what not to ship. Anyone can generate ten options. The value is in reading them and seeing the one that holds, or noticing that none of them do.
The same standard applies to what I build myself. Localization QA started as a standalone tool for checking translated copy against our rules and guidelines. As more people used it, we decided to scale it and it evolved into a Claude skill. Teams can run different checks and get the report directly in their Claude Code session. When content designers update the guidelines or the content design system, those changes carry through to the skill. The results so far have been encouraging. The useful decision was not to keep the tool because it worked. It was to change its form so people could use it more easily.
Judgment does not show in the product
Judgment is invisible in the product itself. What ships shows what was made; it does not show the options rejected or why one version held and the rest did not. The better the result, the more obvious it looks, as if there had never been another way. Message Explorer tells you what it does today. It says nothing about what I decided it should not do.
The product cannot carry any of that. The judgment behind it survives in the people who watched you make those calls. When anyone can generate, that is the one thing nobody can generate for you.