The prompt engineer is fading and that suits me fine

On July 21, Anthropic posted a screenshot of Claude showing a window that offers to record your screen while you work. Your clicks, your typing, and your voice go to the model, which turns them into a skill it can reuse^[1]. Two small labels at the bottom of the image tell the rest of the story, "Recorded demonstration, 23.4s" and "Saved as /file-expenses." Twenty-three seconds of video to learn a job, or at least the part of a job nobody has ever asked for credit on.

It's open to Pro, Max, and Team subscribers in the desktop app^[1:1]. Until now, teaching Claude a piece of know-how meant hand-writing a markdown file^[2] with its instructions and its header, a chore that put off everyone except the people who enjoy that sort of thing, myself included. From here you do the task once, narrate it as you go, and the machine writes the file. That small shift says more than most model announcements, because it stops chasing raw capability and goes after the gap between the technology and everyone who owns neither a terminal nor the patience for one.

The agents are coming down from the terminal

Claude Code, the tool that reorganized a good part of my working day and a fair slice of my budget, launched in February 2025 in the least welcoming form available, a command line. Anthropic only slid it into a desktop app with file trees and progress panes in November^[3], then built Cowork in January 2026 for people whose job isn't writing code, and on July 7 moved the whole thing to the web and the phone, where tasks keep running after you close the laptop^[4]. These are the same diligent agents that pushed most web traffic over to the machines, and now they're knocking on the door of the rest of the office.

The numbers Anthropic released along the way are worth a look. Across 1.2 million sessions observed in May at more than 600,000 organizations, software development accounts for just 8.7% of use, far behind business operations at 33.4% and content production at 16.4%^[4:1]. A tool born in a developer's terminal now spends most of its day in spreadsheets and status reports, which is probably the best news of the year for anyone who thought AI belonged to a priesthood.

Anthropic isn't alone on this road. OpenAI shipped a twin feature in Codex on June 18, where a recorded demo also becomes a replayable skill, on macOS only and outside the European Economic Area, which leaves those of us in France to admire the launch video^[5]. Google opened Workspace Studio, where you describe in ordinary language what you want automated in Gmail or Drive and Gemini builds the agent, with no code and no syntax to memorize^[6]. Mistral renamed Le Chat to Vibe and gave it skills of its own for recurring work^[7]. When four rivals run the same direction inside six weeks, it stops being a hunch and becomes a shift.

One detail makes that shift legible. Anthropic published the skills specification as an open standard in December 2025, and OpenAI, Google, Microsoft, and GitHub adopted it within weeks^[8]. The file Claude writes by watching you work is the same file I write by hand for this blog. We produce the same object. Claude takes twenty-three seconds and I take an evening.

The prompt engineer had a short career

Think back to spring 2023, when the press announced the arrival of the prompt engineer, the specialist whose job would be phrasing everyone else's requests. You can still measure the enthusiasm in Indeed's archives, where searches for the role climbed from two per million in January 2023 to a hundred and forty-four by April, then fell back to twenty or thirty and settled there^[9]. Three years on, Microsoft asked 31,000 workers across 31 countries which roles their employers planned to add, and prompt engineer landed second from last^[10].

Two things happened at once. Models learned to guess the intent behind a clumsy instruction, so yesterday's magic phrasing stopped paying, and the tools absorbed the technique until it disappeared from view, which screen recording finishes off because there's now literally nothing to write. The skill itself survives, and only the job title evaporates. Framing a request well still helps, the way typing fast still helps, and nobody has built a profession on that. The people who lost out weren't the ones who wrote bad instructions, they were the ones who could do nothing else.

I can't bring myself to mourn it. A job that consists of speaking correctly to a machine is a patch over a badly built machine, and its existence was always an admission of failure somewhere upstream. Losing it frees up time for better questions.

What no model has ever wanted to build

This is where the subject stops being theoretical for me. I'm a consultant and project manager in tech, not a developer, and what I like about the work comes down to one move, taking a badly expressed need and turning it into something that actually helps. I could read code without producing it, so my ideas used to die at the exact point where someone had to build them.

They don't die anymore. Meetingtime tracks speaking time and topics so a meeting ends when it's supposed to, Couplecards deals out activity cards when a couple's routine starts to set, and Taster files my films, games, and books into a filterable shelf. Three apps I use for real, each born from a specific irritation, none of which would exist if the tooling had stayed with the engineers. I've since taken to building my presentations in HTML rather than slides, because once you've made something yourself you rarely stop at the first attempt.

No model would have come up with Meetingtime, for the plain reason that no model has ever sat through forty minutes of a meeting going sideways. Machines recombine learned patterns, and they've been better at that than us for a while, while wanting something comes out of a lived annoyance, a whim, or a memory. That's the shift, and I find it cheering. The value no longer sits in how you phrase the request. It sits in having something worth asking for. At the very moment everyone is wondering what will replace them, we're being handed the tools to answer.

As long as the brain stays plugged in

None of which means easy should be confused with excused from thinking, and the early research counsels caution. A team at the MIT Media Lab had fifty-four volunteers write essays under EEG, some with ChatGPT, some with a search engine, some with nothing. The ChatGPT group showed the weakest brain connectivity of the three, struggled to quote their own text, and felt less ownership of what they'd just written, an effect the researchers call "cognitive debt"^[11]. The study is still awaiting peer review, so take it for what it is, an early signal rather than a verdict.

The second signal comes from Microsoft Research and Carnegie Mellon, who surveyed 319 knowledge workers about 936 real tasks done with AI. Six in ten admit to less critical thinking than before, mostly on routine work, and the sharpest correlation fits in a sentence, the more you trust the machine the less you think, while trusting yourself does the opposite^[12]. We hand over our judgment most readily when we assume we couldn't do better.

So the spread of advanced features cuts both ways. It gives back a serious amount of time, and only we get to decide whether that time goes back into thinking or into not thinking at all. Nobody will settle it for us, least of all the company selling the software, whose business is getting us to use more of it.

Which leaves the question I've been turning over since that screenshot, and it isn't a technical one. If Claude can learn my movements from twenty-three seconds of video, what do I intend to do with the hours it gives back? The answer won't come from Anthropic, or OpenAI, or any screenshot on a pale green background. It takes exactly the effort the machine saves us, an irony I find rather well made.