I decided to skeptically play with Claude Code recently.

[recursive@hachyderm.io]

@EndlessMason I saw shellcheck catching on starting about 10 years ago and honestly feel like it's made people a bit complacent

[recursive@hachyderm.io]

"Write more code" is just a terrible incentive structure

[ret@furry.engineer]

@recursive despise it. delete more code. That’s the key to success.

[recursive@hachyderm.io]

I still have extremely mixed feelings about these tools, but a friend actually convinced me that they can be useful sometimes

I worry most about imposing cognitive load through shitloads of code to review

Well that and management misunderstanding of the whole mess leading us to unmaintainability

[jesstheunstill@infosec.exchange]

@recursive
Boo, hiss!
You can take my shell scripts from my cold dead hands!
...
Yes I know I should just do shit in python or whatever, but a simple shell script to do a little curl, jq, grep, sed, etc. without having to faf around with actual code

[recursive@hachyderm.io]

@JessTheUnstill IME the problem is that other languages tend to fail at the expressivity we are accustomed to in shell script. They make "run a command" a big deal and they don't give good syntactic sugar around it

And a lot of them are languages which are historically loathe to introduce clever syntax (e.g. python)

[dannyman@sfba.social]

@recursive I was talking with a colleague about AI over the weekend. Mentioned that some developers started sending in a lot more code to review. And they buffered that by having AI do the initial PRs. Kind of like an advanced linter. Took some of his back and forth out of the loop. Because the the AI based engineers could get AI feedback on their AI code. Then when that was all polished up you can bring it to the senior human.

[mnl@hachyderm.io]

@recursive reframe their use as "how can I use this tool to assist me create higher quality software. Some of it might be writing the code with the LLM, to me it's more about building tools to reduce cognitive load (thus allowing more cognition to be focused on the real issues), and being able to iterate on many prototypes / try out many different things quickly, code which will not make it into production or is even supposed to survive for very long, in order to map out the design space which will allow "simple" (as in Simplicity with a capital S) solutions to be found.

One of my favourite prompts which costs almost nothing: "design an elegant API to do X" (plug in your problem, point at your existing codebase, etc...), and then go "oh yeah that's pretty neat, build" or "nah i don't like that, do it more like Y". In the worst case you'll build the API implementation yourself, but by virtue of being LLM generated, the API is usually pretty consistent, non-surprising, using well-known patterns.

[jesstheunstill@infosec.exchange]

@recursive I got good at shell back in my Linux Admin days where who knew what tools might be on any given node, and they didn't have internet connections to pull down dependencies. Hell, plenty didn't even have python installed. So it's a lot of living off the land knowledge I got used to.

[jesstheunstill@infosec.exchange]

@recursive command line and file operations are also just way more headaches than they're worth. I'm sure if I did it enough I'd memorize their syntax, but even all these years later, still nobody can keep up with the speed of grep. It's a magic program.