Phoenix.new – Remote AI Runtime for Phoenix

Ah man I'm really happy to see this and excited to try it out.

As an Elixir enthusiast I've been worried that Elixir would fall behind because the LLMs don't write it as well as they write bigger languages like Python/JS. So I'm really glad to see such active effort to rectify this problem.

We're in safe hands.

This last few weeks I've been going hard on LLMs to put together a new prototype project. I've exclusively been using Claude Sonnet 3.7 within Zed (via github copilot) and it' fantastic.

From time to time it tries to do something a little old-school, but nothing significant really. It's very capable at spitting out entire new features, even in liveview.

Over all the experience has very productive, and at least on-par with my recent work on similar sized python and nextjs applications.

I think because I'm using mostly common and well understood packages it has a big leg up. I also made sure to initialise the phoenix project myself to start with so it didn't try to go off on some weird direction.

Assuming you are on the $20 Zed plan? Has the 500 prompts/mo been sufficient for you? I'm debating between the Zed and Claude $20 plans-- no doubt I'd get better value from Zed's?

I have used Zed's plane with Claude and also Claude Code. They are very different experiences. Zed's agent work is very much a set it, go away, review, give some tips to it, iterate. As long as you use the Sonnet and absolutely avoid the burn mode (formerly max mode), it should do a lot of work for you. The main limitation I hit is the context window. As the codebase gets larger, it takes more context for it to get going and then it tends to have a hard time finishing. I find that about 4 prompts works for a feature that would take me a few hours to code.

For Claude Code, the limit is reset every 5 hours so if you hit it, you rest a bit. Not that big a deal to me. But the way it works I find much more stressful. It is reviewing just about everything it is doing. It is step-by-step. Some of it you can just say Yes, do it without permission, but it likes to run shell commands and for obvious reasons arbitrary shell commands need your explicit Yes for each run. This is probably a great flow if you want a lot of control in what it is doing. And the ability to intercede and redirect it is great. But if you want more of a "I just want to get the result and minimize my time and effort" then Zed is probably better for that.

I am also experimenting with OpenAI's codex which is yet a different experience. There it runs on repos and pull requests. I have no idea what their rate/limit stuff will be. I have just started working with it.

Of the three, disregarding cost, I like Zed's experience the best. I also think they are the most transparent. Just make sure never to use the burn mode. That really burns through the credits very quickly for no real discernible reason. But I think it is also limited to either small codebases or prompts that limit what the agent is going through to get up to speed due to the context window being about 120k (it is not 200k as the view seems to suggest).