[cadamsau]

One hundred per cent this.

LLM pair programming is unbelievably fun, satisfying, and productive. Why type out the code when you can instead watch it being typed while thinking of and typing out/speaking the next thing you want.

For those who enjoy typing, you could try to get a job dictating letters for lawyers, but something tells me that’s on the way out too.

[not_kurt_godel]

My experience is it's been unbelievably fun until I actually try to run what it writes and/or add some non-trivial functionality. At that point it becomes unbelievably un-fun and frustrating as the LLM insists on producing code that doesn't give the outputs it says it does.

[adamredwoods]

I've never found "LLM pair-programming" to be fun. I enjoy engaging my brain and coding on its own. Co-pilot and it's suggestions after a point become distracting. I'm sure there's are several use cases, but for me it's a tool that sometimes gets in the way (I usually turn off suggestions).

[wyclif]

What do you prefer to use for LLM pair programming?

[richardw]

Claude 70%. ChatGPT o1 for anything that needs more iterations, Cursor for local autocomplete, tested Gemini for concepts and it seemed solid. Replit when I want it to generate everything, from setting up a DB etc for any quick projects. But it’s a bit annoying and drives into a ditch a lot.

I honestly have to keep a tight rein on them all, so I usually ask for concepts first with no code, and need to iterate or start again a few times to get what I need. Get clear instructions, then generate. Drag in context, tight reins on changes I want. Make minor changes rather than wholesale.

Tricks I use. “Do you have any questions?” And “tell me what you want to do first.” Trigger it into the right latent space first, get the right neurons firing. Also “how else could I do this”. It’ll sometimes choose bad algo’s so you need to know your DSA, and it loves to overcomplicate. Tight reins :)

Claude’s personality is great. Just wants to help.

All work best on common languages and libraries. Edge cases or new versions get them confused. But you can paste in a new api and it’ll code against that perfectly.

I also use the API’s a lot, from cheap to pricy depending on task. Lots of data extraction, classifying. I got a (pricier model) data cleaner working on other data generated by a cheaper model, asking it to check eg 20 rows in each batch for consistency. Did a great job.

[traverseda]

Copilot and aider-chat

[cadamsau]

Claude and pasting code back and forth for simple things. But I’d like to try Claude with a few MCP servers so it can directly modify a repo.

But lately Cursor. It’s just so effortless.

[nidnogg]

Yeah I had my fair share of pride around typing super fast back in college, but the algorithms were super annoying to think through.

Nowadays I get wayyy more of a kick typing the most efficient Lego prompts in Claude.