I’m a web developer and each week I am more positively surprised what Claude is capable of. I feed it small tasks, and review every line of code it produces.
We are working on a closed source framework so Claude has no specific knowledge of it. With the use of AGENTS.md and some specific skill-files, it has mastered our codebase as good as the average senior developer.
Using Claude, we can deliver way more polished features for our customer in less time than manual coding. This is a win-win for all of us: the customer is more happy, and so we’re more happy. It lessens our workload and we can spend more time on really helping the customer and polishing new features.
Prototyping is really great as well! You can let Claude create a feature which you can actually test before you decide to go further in that direction, or just revert all changes. If you coded it yourself, you’ll probably hate to revert the changes, but now it just the press of a button.
Then there’s the environmental issues. I like to think that our LLM usage is way more acceptable than the AI generation of cat pictures, but still bad. Sorry.
I can agree on everything you said.
The thing i want to point out is - It’s always easier to rebel about something somebody else created, than to create something admirable yourself.
Like you, I like to code. But I now think I like to code to generate useful and meaningful stuff for our paying customers. Using (supervised) Claude I can create even more cool stuff in less time (read: money) and this makes me and our clients more happy.
I also hit less tiresome roadblocks and Claude also showed me other ways (often better) to solve issues.
Using Claude to translate stuff also saved me a lot of hours of grunt work.
A lot of the discussion is around LLM generated code within Ardour, but the way I see LLM code generation being used in my industry is as a very sophisticated variant of Lint, coverity, static analysis, etc. along with code generation for automated test suites.
The software and firmware developers at my day job are not allowed to use LLM to generate any source code for shipping products due to the copyright situation explained by Paul, but internal test tools are not distributed, so copyright is not a concern, and of course analysis of code results in descriptive output of potential quality problems, not code, so also not of concern.
It seems like a similar use could benefit Ardour in hopefully finding subtle problems before release.
