It works great for a number of reasons

- it’s functional making it much easier to reason about for LLMs (eg no side effects)

- it’s compiled (including the html template), so the LLM get instant feedback if something is off and can fix it quicker

- it’s more concise than other frameworks (language, framework, no front/backend split) and consequently you hit the max token limit much less frequently

- it has excellent tailwind support

I told a business partner to just use claude for certain tasks a year ago, and it failed miserably in python where it succeeded in elixir/phoenix liveview. This was a regular occurrence. LLMs have obviously progressed since, but the principle still stands that they work much better in elixir than python.

So in short: it’s great and arguably better than most other languages and frameworks

I’d like some confirmation if possible, but my gut says the general internet (and therefore the training data) probably has a lot of sloppy Python code along with a lot of sloppy writing about Python whereas Phoenix has majority good code and well-reasoned writing

I'd argue that it's one of the best, especially with Tidewave:

https://dashbit.co/blog/why-elixir-best-language-for-ai

LLM use is very solid here. I have been shipping some fairly cool stuff for different clients, just like if I was managing a team of 3 or 4 people.

The stable nature of the language, and the static/immutable nature of constructs, presumably really helps moving safely here.

I've built several production quality, mid-sized apps in Phoenix/Elixir over the last year using Claude Code (CC). In my experience, it's great, I had better results with this stack than I had with Django when using purely CC. It's token efficient (because of code generator), has great static analysis, and terse (because functional). Phoenix even generates an Agents.md The apps themselves are highly scalable and look great.

Checkout https://github.com/oliver-kriska/claude-elixir-phoenix (not me).

Pretty good as others have mentioned. Even more so when one starts adding packages like credo, ex_slop, and boundary to run deterministically and steer the LLM.

Don’t like a pattern that LLM introduced? Write a custom credo rule and the problem doesn’t come back.

It’s exceptional. We’ve improved throughput on new features by about 2x and sentry errors are down. Our codebase is 2-300k loc and pretty old in some places, I’ve been really impressed with just sonnet.

Worth checking out...

https://phoenix.new/

They were pretty good a year ago, so I am sure it is even better than that now.