Prompting Is Basically Over? Anthropic's Cat Wu Builds a Self-Improving System from an Empty Terminal in 10 Minutes

Solen shared a live demo from Cat Wu. Solen's tweet calls her an Anthropic engineer, while the on-screen text in the video says Cat Wu Product Manager. In the demo, she made one judgment: “Prompting is basically over.” She said 95% of engineers on the team are already running self-improving loops, and the next step is graphs. Then, in 10 minutes, she built the Claude Code setup live from an empty terminal, following the path Empty Terminal → Agents → Loops → Graphs → Self-Improving System.

The focus of the demo is the process. Many people still understand AI programming as writing prompts, asking one question and getting one answer. She breaks the process down into a system: an agent connects to tools, loops repeatedly execute and check, graphs organize multiple steps, and finally the system improves based on results. Solen's tweet puts it more directly: prompting is the old way, and doing systems engineering around the model is the new way. This judgment has a promotional tone, but the direction is not new. Writing prompts on their own is still useful, but it is hard to make multi-step tasks reliable with prompts alone.

For readers without a development background, loops can be understood as letting an agent repeatedly execute, check, and execute again. Graphs can be understood as mapping multiple steps and dependencies into a workflow, so that everything does not get crammed into a single straight line. One reply mentioned that approving only the final step gives an agent graph more room to experiment. Another reply warned that agent loops are one of the productivity killers. Without evaluation and approval stages, loops can easily spin their wheels. This reminder is more useful than the terminology itself.

Solen says this demo is worth more than most $1,000 agent engineering courses. The empty-terminal starting point in the video is more instructive than a screenshot of the finished product; one reply liked this because the setup is no longer a secret handshake. The tweet also links to Hanako's Graph Engineering guide, recommending that people bookmark it, watch the video, and then read the guide. Guide link: http://x.com/i/article/2091485629452304385

Knowing how to write prompts is still a foundational skill. System-level things such as loops, graphs, evaluation, and approval determine whether an agent can deliver reliably. This demo sets out the threshold: whether you can build an agent system starting from an empty terminal is closer to real work than whether you can write a pretty one-line prompt.