The code I generate with LLMs is clean and looks as good if not better than what I'd write by hand.

I don't get all the prompt vibe coding going around. I don't use prompts to generate code.

I use "tab-tab" auto complete to speed through refactorings and adding new fields / plumbing.

It's easily a 3x productivity gain. On a good day it might be 10x.

It gets me through boring tedium. It gets strings and method names right for languages that aren't statically typed. For languages that are statically typed, it's still better than the best IDE AST understanding.

It won't replace the design and engineering work I do to scope out active-active systems of record, but it'll help me when time comes to build.

I use tab auto complete, and i think it's a 5% productivity gain. On a good day, maybe 10%. I haven't put much effort into optimizing the setup or learning advanced usage patterns or anything. I'm using stock copilot, provided by my employer. If I had to pay for it, I wouldn't be using it, as it doesn't justify the cost.

Really, what are you making that a 5% increase in productivity doesn’t justify a Copilot subscription?

That's not a rigorously measured number.

The 5% is an increase in straight-ahead code speed. I spend a small fraction of my time typing code. Smaller than I'd like.

And it very well might be an economically rational subscription. For me personally, I'm subscription averse based on the overhead of remembering that I have a subscription and managing it.

Prompts are worth mastering. AI autocomplete is better than older autocomplete systems but of course it only works based on what you started to type.

Prompts are especially good for building a new template of structure for a new code module or basic boilerplate for some of the more verbose environments. eg. Android Java programming can be a mess, huge amounts of code for something simple like an efficient scrolling view. AI takes care of this - it's obvious code, no thought, but it's still over 100 lines scattered in XML (the view definitions), resources, and in multiple Java files.

Do you really want to be copying boilerplate like this across to many different files? Prompts that are well integrated to the IDE (they give a diff to add the code) are great (also old style Android before Jetpack sucked) https://stackoverflow.com/questions/40584424/simple-android-...

> For languages that are statically typed, it's still better than the best IDE AST understanding.

This is emphatically NOT my experience with a large C++ codebase.

I can't attest to C++, but we've got a large Rust monorepo, and it's magical.

It expands match blocks against highly complex enums from different crates, then tab completes test cases after I write the first one. Sometimes even before that.

We may be at different levels of "large" (and "gnarly") - this code-base has existed in some form since 1985, through various automated translations Pascal -> C -> C++.

Just by virtue of Rust being relatively short-lived I would guess that your code base is modular enough to live inside reasonable context limits, and written following mostly standard practice.

One of the main files I work on is ~40k lines of code, and one of the main proprietary API headers I consume is ~40k lines of code.

My attempts at getting the models available to Copilot to author functions for me have often failed spectacularly - as in I can't even get it to generate edits at prescribed places in the source code, follow examples from prescribed places. And the hallucination issue is EXTREME when trying to use the big C API I alluded to.

That said Claude Code (which I don't have access to at work) has been pretty impressive (although not what I would call "magical") on personal C++ projects. I don't have Opus, though.