John Elm Labs

@thdxr Why can’t you just have an agent doing this ? Why a person?

are there people out there who just want to refactor every day? just wake up and find the worst code and just chip away at it and clean it up wake up the next day do it again, infinitely improving things with zero external impact?

I wanna see the global Git commit graph overlaid on Claude’s uptime, I bet it’s a 1:1 correlation

My (unproven) hypothesis would be that you could write the current in-flight tasks to disk and read them back on start up to kick them back off Things you lose with this would be scheduled / future / cron, but that’s not my goal I don’t need all the connections & polling for simple job retries. Database is always the bottleneck in prod. Elixir systems I’ve worked on. I think a stupid simple retry would serve most use cases Oban is great and when the needs dictate the move to it, I’d move to it. It just seems there’s room for a smaller, simpler starting point

@JohnElmLabs I’d say that for very simple retries an arrangement of Tasks would do it. But then, you’ll end up stumbling on things like “restart survival”, logging, error capturing, etc., etc. — and then you’ll find yourself writing your own Oban. We call it “reinvent the wheel.”

I’ve always felt like for simple job / function retry there was a lighter dep. you could bring than Oban I think I can finally vibe code it now All OTP primitives and file system….here goes nothin

I don't even use my editor anymore. It's just Claude and the GitHub diff window

@sean_moriarity Have you seen openai.com/index/buy-it-i… ? developers.openai.com/commerce/

I feel like there is a layer missing right now in the development flow with AI. We’re at the point where I want agents to authenticate and make purchase decisions on my behalf (with my approval). Why can’t I just give them a credit card and say: “find me an API that does X, ask for permission, and give me a key” It’s easy to develop against these APIs, but there’s still friction in researching, setting up billing, generating credentials, etc.

@josevalim Wow! Congratulations, and best of luck on the journey ahead

Almost 5 months since launch, Tidewave has crossed 100k ARR! 🎊 The best part is that we can invest all profits back into the product, since we are not reselling/subsidizing tokens. Thanks to everyone who is using it and for all of the feedback!

Codebase architecture is a skill that is never taught and yet it's now more valuable than ever.

AI is amazing at writing code. Amazing. It's awful, and I mean awful, at organizing code It makes sense when you know AI is just a _really good_ next word predictor

Elixir Radar issue 491 is out! 📣 You can read it here: buff.ly/alySdZq This issue comes with content from @JohnElmLabs @AtomikSnowman @mrpopov_com @jaguar0ndi , Stuart Page and Peter Ullrich. Thank you! #ElixirLang #MyElixirStatus

@JokingRajat What if you folded me into the closed beta 🫣

@JohnElmLabs Will be opening up the closed beta soon!

Update #10 on Diff - a fast and local PR reviewer app Quickly change the base branch of your PR in the app.

Anybody gotten a decent system to "compose" claude slash commands? Something like "do feature [x]. /test /format /commit /open-pr #MyElixirStatus

TOON is premature optimization

@bcardarella @ziglang @elixirlang Claude Code without Tidewave is a nightmare

I have been agentic coding in @ziglang for over two months and today switched over to @elixirlang to build a basic Phoenix app. I can say definitively that with the same setup the LLM is *way* dumber writing Elixir than Zig. I suspect the following: 1. I'm writing a phoenix app and not using tidewave so insight to the server environment is more difficult 2. my Zig work didn't require any outside dependencies so outside of the Zig stdlib the entire API surface is available 3. Zig's typesytem I have some ideas how to improve this, going to experiment.

@mmmykolas @form Yes, checkbox + label can absolutely be styled to work with strictly the form In practice, I find few devs I've worked with know you can style labels like that ;)

I've read your post. You can still simplify this. You don't need javascript. <.form for={@form} phx-change="validate" phx-submit="save"> <.input type="checkbox" field={@form[:has_referral_code]} class="peer sr-only" /> @form[:has_referral_code].id} class="inline-block cursor-pointer px-3 py-2 bg-gray-200 rounded" > I have a referral code <.input field={@form[:referral_code]} label="Referral Code" class="hidden peer-checked:block mt-3" />

The Top 3 LiveView Form Mistakes (and how to fix them) Prevent poor UX, brittle systems, and impossible states by using these 3 strategies. johnelmlabs.com/posts/top-3-li… #MyElixirStatus

HexDocs over swagger for public REST api documentation? Good Idea? bad idea? #MyElixirStatus

Claude fails to add some code comments and loses its mind 😭

@bcardarella I think you'd benefit from github.com/chrismccord/web -- Converts html to markdown & can execute JS, which is good for stuff like Apple's docs. Apple docs can also be read by sosumi.ai MCP server

LLM Codegen against well written specs is kind of magical. However, there are some optimizations I've had to do: 1. fetch original spec document (usually in HTML) 2. pandoc convert to markdown 3. ask LLM to analyze and remove anything from the document that is unnecessary and optimize for tokenization Without losing spec fidelity: DOM: 760k tokens to 62k tokens HTML: 3.7M tokens to 750k