Building a self-improving system from an empty terminal in ten minutes: Anthropic engineer says prompting is over

In a video, Anthropic's Cat Wu stands in front of a screen and offers a judgment: prompting is basically over. She mentions that 95% of engineers are already running self-improvement loops, and the next step is graphs.

Then, in 10 minutes, she builds a complete Claude Code setup from an empty terminal. The path is clear: empty terminal → Agents → Loops → Graphs → Self-Improving System.

Let's clarify a few terms first. Prompting is writing prompts, with humans figuring out how to ask. Agents let the model call tools and take actions. Loops let the model repeatedly run the same process and correct based on results. Graphs connect multiple agents or steps into a directed graph, so when something fails you can locate which segment it was. A Self-Improving System takes outputs back, evaluates them, and feeds them into the process again.

The point of this chain is not the slogan 'Prompting is dead.' What it really compresses is setup time. Previously, building a working agent workflow might take an afternoon of configuring the environment, writing scripts, and trial and error. The video compresses this to 10 minutes, which is useful for people already familiar with Claude Code.

The comments section is livelier than the video. Some point out that the video is old and mentions Opus 4 and Sonnet 4. Some ask how long it would take an ordinary person to get from an empty terminal to this setup. Someone says outright that first you need unlimited tokens. Another says they built a similar system a year and a half ago and there's nothing new.

One reply gets upvoted: when she says prompting is over, the title on screen is still Product Manager, and the engineers she knows are still begging ChatGPT to write emails. Graphs and such are for people who have forgotten what it feels like to struggle with a CSV for an hour.

That's a bit harsh, but it points to a fact. The video shows an ideal path; in reality, token costs, task stability, and maintenance time will all get in the way. Self-improving loops sound advanced, but running them may first burn through a pile of API call fees.

What's more interesting is the term 'graphs.' It doesn't make the model smarter; it maps out multi-step processes so each step is observable, replaceable, and rollback-able. For readers with a development background, this is closer to workflow orchestration. For ordinary users, its significance is that in the future they may not have to watch every step themselves—but only if someone builds the graph first.

The video is old, the discussion is new, and this mismatch itself is more informative than the demo. So the video is worth watching, but it doesn't need to be a source of anxiety. First, find a task in your own hands that you've repeated more than three times and see whether it can be broken into a few steps. If it can be broken down, then consider a loop. If it can't be broken down, prompting is still far from over.

Original link: https://x.com/halvenruns/status/2116625585854431581