I’ve been using Claude Code quite a bit this week. It is amazing at a lot of the basic infrastructure that is part of science research. Examples:
- producing quality rsync scripts for repeatable downloads
- setting up a testing harness for some code I am writing
- debugging moderately complex issues and suggesting a fix
- setting up a website using Quarto (no previous experience)
None of these are things that I could not do. However, Claude Code gets it done much faster. Also, the code quality is higher for the time taken. Take the rsync scripts - I would have made something that worked in about the same time that Claude used to make something that worked, was extensible, had command line arguments and a guide on how to use it. AND it showed me a new rsync switch: --info=progress2, which shows you what percentage has been transferred and is awesome. The downsides to using it are a) you need to pay for it, and b) how can you build understanding?
The last point is an interesting one – I have spent many hours writing bash scripts poorly. I find it impossible to memorize the slightly arcane syntax, and generally have to look up how to do simple comparisons every time. Clearly, I have had ample opportunity to learn how to do it, yet have never devoted the time – every time I wrote a bash script and it grows beyond a certain size, I rewrote it in Python. So am I losing out by not building this understanding? Probably not in this case. But if I was starting out now, would I be able to develop a feel for what was possible, and how to approach certain problems? I’m not sure.