The – new – heavy plough

What Claude Code means for the future of work

On a flight from Cape Town to Joburg at the end of last year, I sat behind a businesswoman making slides. She was in retail, and spent two hours on a 70-page deck. (She was sitting on the opposite side of the aisle, one row ahead. It was difficult not to look.) Let me be more precise: when she opened her laptop, the deck was already finished. She spent the flight nudging it into shape, so that the images (lots of photos of Ricoffy and All Gold – South Africans will know) sat just so.

That cannot be a nice job. I could see the struggle: dragging a photo with a trackpad, then dragging it back, then adjusting the text box that had moved when the photo moved. An inch this way, an inch back. She hated it, and I hated her for forcing me to watch.

Then Claude Code arrives. Two weeks ago, while I was preparing my slides for the new semester, I asked it whether it could update all my LaTeX slides with a new style in one pass, and make sure the bibliography still compiled. I added a few more requests without much thought: standardise the tables, run a language check, fix typos, and add a translation as well. What would have taken me days took me as long as writing that paragraph.

The temptation is to describe this as speed. Speed matters, but it is the least interesting part. What changes is what you attempt in the first place.

Here is the econometrician Scott Cunningham, posting last week:

I’ve been using Claude Code since mid-November. I remember the first time — I used it to fuzzy impute gender and race in a dataset of names. I thought I was just looking at another chatbot. But I kept using it, kept taking more steps, and kept being stunned at what I was doing with it. By late November, I’d used it on a project that was sufficiently hard, and that was when I knew. The speed, the range of tasks, the quality — it did not have an easily discernible upper bound. More work, less time, better work — but also different work. Tasks I wouldn’t have attempted because the execution cost was too high suddenly became feasible. Some combination of more, faster, better, and new at all times. It was, for me, the first real shift in marginal product per unit of time I’d experienced, even counting ChatGPT. I felt a genuine urgency to tell every applied social scientist who would listen: You have to try Claude Code. You have to trust me.

You have to trust him. I will have much more to say about Claude soon, but for now let me describe it in the only way I know: through an economic history analogy…