They are building exactly what you described and this is their architectural solution to ensuring their YOLO agents do not nuke their customers code/documents/databases by sandboxing everything in the workspace — the git checkout the agent is working on, plus whatever's needed to run commands against it (compilers, package managers, etc.).
A well written post on a well-thought out product.
> Eventually, I want Ruin to be a native Mac and iOS app.
Whenever I've had this idea for other projects the first concept I like to play around with is a Raycast extension, it allows you to experiment with how you can leverage spotlight search and Mac commands to interact with your product, especially for the Quick Capture functionality (I see you've mentioned it toward the end).
I like the concept, is there a wizard type thing for how I can port over a lot of my existing .md files to the "ruin style .md files" with all the notation? Perhaps something I can give Claude to essentially onboard my files to be viewed and interfaced with effectively through ruin?
Good that they prefaced it with this "Claude can't replace taste or imagination". I think this is a solid step in the right direction and the more tools Claude has access to the better (more surface area == faster iteration == faster tinkering).
I've worked with Claude in many creative capacities and it's issue is that despite it being able to see if you ask it to draw something (using ascii for example) it will fail, if you ask it to iterate on that drawing it will continue to fail and not get any closer to the target then complain about this.
I've felt that these models struggle with anything that cannot be decomposed into primitives and their architecture is too greedy and favours the obvious, autoregressive generation so it will converge to the modal answer. So unless they have enhanced the models in some creative sense I fail to see how this is anything other than giving Claude a bunch of documentation/MCP servers/APIs/CLI tools (which already existed) and making an announcement out of it.
My point: FREE the models, unchain them and let's see what they are actually capable of, also put some damn demos in the announcement post???
Seems like outages are increasingly more frequent nowadays. Obviously, this is not the best state of affairs, and developers should not be limited by services. In the meantime I've been experimenting with building third spaces for people to chill while they wait for the services they are dependent on to go back up.
The first one I've built is a little ASCII hangout for Claude @ https://clawdpenguin.com but threads like this make me want to build it for Github too.
Claude's down again! So while it's down, come chill in my new hangout space: https://clawdpenguin.com
It's a little ASCII world you drop into — roll a companion (18 species, gacha rarity, 1% shiny), wander around, chat with whoever else got kicked out of their Claude tab. Built on a Cloudflare Worker + Durable Object for the multiplayer state, spatial grid for nearby player lookups, plain canvas rendering with no build step on the client. Sprites are the buddy sprites pulled straight from the CC source. There's a live Claude status board in the corner so you know when it's time to get back to work.
Your app will not succeed if users have to login and create an account to test it out. Have a free tier with rate limiting at least, absorb some of the cost if you truly believe in promoting this product.
I think we're arguing two different points. You're arguing about implementation, AI is great at this given the existing defaults and the right prompting. AI was trained on 30+ years of accessibility standards that a minority of great humans fought to establish as a familiar practice.
I'm arguing about invention. It is extremely unlikely that AI will be the one to invent the next accessibility paradigm, because that requires deviating from the training distribution, which it CAN'T DO.
I'm also arguing that this homogeneity in design will lead to an atrophy in inventive, unique and original thinking.
I should have mentioned accessibility. It supports my argument more than yours. Accessibility like captions, voice, keyboard nav, dark mode are all a deviation from the norm by a minority (something AI is completely incapable of doing) and a fight against familiarity which now serves as a great benefit to the majority.
You're arguing for familiarity in tactful design, while I agree that for most users this is a good thing, repeatability of existing patterns does create that immediate familiarity.
HOWEVER, that familiarity is only a virtue because someone, once, deviated hard enough that their deviation became the new familiar. AI can only optimise toward the current snapshot of "familiar". It cannot produce the next one. If designers outsource all their thinking to a model even in tactful design we would never have groundbreaking design concepts like "pull to refresh" or the command palette.
I will stand by the first point unless models start being trained with different objectives instead of RLHF's three objectives: Helpfulness, Harmlessness and Instruction-following
The more I think about it the more this isn't good for design [EDIT], for a few reasons:
- The best design is original, groundbreaking and often counterintuitive. An AI model is incapable of that, it's uninspired, it will absolutely converge to the norm and homogeneity (you see it everywhere now, just scroll on ShowHN and take a look at the UIs) and produce the safest design that appeals to its understanding of the ideal user.
- Good designers will reject this, they prefer to be hands-on and draw from multiple sources of inspiration which is what Figma boards and Canva is good for, also mainly for cross-collaboration. If you've seen how quickly a great design engineer can prototype you'll know that "speed" they advertise in this video is not worth the tradeoff.
- Creatives typically have a very very very high aversion to AI.
- Non-designers will not see a purpose for this tool, basic design can already be done through Claude Code and Claude.ai, I fail to see what this could offer unless they leverage a model that is more creative and unique by default (you can not prompt/context/harness engineer creativity believe me I've tried).
- Design is a lot more than just UI. Tools like this ignore so many other important aspects like: motion, typography, images, weight, whitespace, sound, feel.
I'm not an supporter of blindly adopting the "langs" but langgraph is useful for deterministically reproducable orchestration. Let's say you have a particular data flow that takes an email sends it through an agent for keyword analysis the another agent for embedding then splits to two agents for sentiment analysis and translation - there is where you'd use langgraph in your service. Claude Code is a consumer tool, not production.
Langchain is for model-agnostic composition. Claude Code only uses one interface to hoist its own models so zero need for an abstraction layer.
Langgraph is for multi-agent orchestration as state graphs. This isn't useful for Claude Code as there is no multi-agent chaining. It uses a single coordinator agent that spawns subagents on demand. Basically too dynamic to constrain to state graphs.
> mental model of the hiring process is that the company is probably going to sabatage your end-game interviews
I definitely agree and it is not a mental model that I carry into any interview, I have good intentions and I'm super friendly! This was only a tiny (disillusioned) post-interview reflection. I would say most interviews especially with engineers have gone well but there has absolutely been a vibe shift in the past year.
You can tell teams are a lot more risk averse when it comes to hiring. The promise of a fabled 10x engineer on the horizon paired with SWE automation devaluing existing talent has meant they will make you jump through 10 more loops and even then the decision is scrutinised. Understandably hiring is an expensive process (both successful and unsuccessful).
> Most employers will want some assurance that you are serious about the position you're applying for.
This is also a reflection of the job market. If it was balanced this notion would not exist. It's become a game of numbers, automated screening + AI has meant candidates need to send out 100s of application often with automation on their end too. On the other side every job likely receives 1000s of applications especially with stupid things like "L*nkedIn Easy Apply". Me personally, I would not apply for a role I am not committed to taking and I especially would not have gone through FOUR stages for fun, the first interview should be plenty screening for both parties!!! Alas.
I appreciate you taking the time to respond and thank you for your well wishes!