Fire under my Ass!

I promised this series almost six months ago, but I'm finally getting this started. Seeing how February is a perfect month, it is apt that this series starts this week. Moreover, this week has been insanely productive, partly because of a new fire lit under my ass and partly because of a great unlock in using AI. Read below to know more, but moving forward, there should be a new update every Saturday or Sunday wherever I am (even tho this one is coming on the ~Tuesday post haha).

Part time role

First things first, I should address my day job and honestly, I have kind of been slacking here. I have excuses, but it is undeniable that I have been slacking. The main problem is that when the project first started about two months ago we got so much done in about two weeks, out of which the last 3-4 days were pivotal. All of that was possible because of using Cursor Agents and completely handing it off. It did end up costing me close to a hundred dollars, but the amount of stuff that was accomplished was pretty crazy. The other issue with the agents was that the code really sucked and while it was somewhat functional and got us through the client demo, it also introduced an immense amount of tech debt. What has since become clear tho is that the Cursor agents were not used effectively or how they are supposed to; ie, no skills, no context, just ask for a feature and let it ship. The features were also perhaps not well broken down.

So, the plan was to be a lot more methodical about the refractor and perhaps the pendulum swung too far the other way and it was very easy to lose motivation quickly. That said discovering how to leverage Claude code effectively presents a new opportunity to complete my pending tasks quickly. I have refactored quite a bit and set up the architecture so I believe Claude code can now handle the completion. That will likely be the primary task for next week. Or at least the first few days depending on how quickly things get done.

VPS Setup

On Monday I finally got around to setting up my VPS. I had been paying Hetzner for about a month or so already and decided it was about time I set it up. Not having a VPS available was really hampering my shipping speed. The main reason for meeting a VPS is because I want to move away from React to start using htmx, and also while at it move away from Vercel. I also need a lightweight analytics tool and I have been traditionally using goatcounter but because of having multiple projects and sites, I have been meaning to self-host for a while. The other thing I've been meaning to self-host is an indexer for all on-chain queries. Almost all of my projects involve interacting with some chain so a shared resource would help spin up new projects faster. Of course I went with Envio's HyperIndex!

The urgency was also directed by needing a landing page for Unhinged so that I could pitch during ETHDenver. I didn't really want to do a quick Vercel site because I also wanted to do some kind of an email-gated deck download. I figured now is good as any time to finally get set up with the VPS. All the VPS setup happened over two days, if that. On Monday it was more 80/20 where I was within Cursor doing most of the things myself and using the Ask tab within Cursor to clarify questions or maybe get some code snippets. No real agent mode usage. And everything was working correctly but making changes was going to be a pain. This is where Claude code really came in clutch.

The Claude stuff happened during HackMoney on Friday. Because of Claude my setup now has different profiles for dev and prod, shared services and proper setups, a way to add projects, and proper convenience scripts to even restart a single app. Sure, there are still some bugs and kinks that need to be figured out, but so far, this is already very accelerating. Now I can actually ship all my projects at a much faster rate, without being locked into Vercel or dealing with React. I really do not like React lol.

Claude Code

This was the major unlock this week, thanks to my friend Trxnt. We have a weekly sync where we chat about design and photography and trends in ai/code, and had previously built a methodology to come up with a full color system for new projects. However, that was based in Figma and was fairly manual and laborious, especially if we wanted to experiment with different styles. We needed a website that could take in the minimal decisions we make and let us visually (methodology designed as such) inspect the result of those decisions. With Figma, even with variables, it was still quite cumbersome; and since Trxnt is focused on building a system to leverage AI better for design related tasks, we figured why not try to use the system?

And oh my god. What an unlock! There were two main things that actually accelerated our workflow,

1. Using speech to text to type in our messages to claude code
2. Using skills for claude code to ensure quality and long term usability

We were on call for about four hours, where we would have actually worked for maybe two and a half, and at the end of it, we got a pretty usable site, with the code correctly formatted (thanks to Vercel's React skill) and well modularized. We went through maybe two iterations of brainstorming before letting it code, and then three iterations after that. The first one was kinda shit, ngl but we gave it a 9-point (~500- word) feedback and the next version was super clean! The long feedback was not typed as well we used speech-to-text (Wispr Flow) which basically meant speaking out loud while clicking through the website output. Insane!

Trxnt had also given Claude chrome privileges (which I am still unsure about lol) but we did notice it noticing bugs during its session and fixing them on the fly! Again, insane! But all of this excitement aside, I think this system is really valuable and I am currently trying move my entire workflow to leverage this system. This will also likely be the base (somewhat) for khaaliDimaag, though I am not entirely sure how exactly. Still, at the very least this system can help me catch up on my [ever increasing] backlog! & perhaps even start chipping away at all the bookmarks and saves and downloads to to-read folders.

HackMoney 2026

Finally, the hackathon! I have been hacking with ETHGlobal for almost five years (damn) now! & I still don't really have the finalist tag! Something, I really need to fix this year. Sure, I have won a bunch of sponsor bounties, but I would very much like to get their finalist pack. One of the biggest problems in how I do hackathons is treat is as sprints for my varied projects, which causes two problems: (i) the scope is often too ambitious for the duration of the hackathon; (ii) aligning with the sponsors ends up diluting the vision for my project, and also feels "half assed". So, I basically get the worst of both the worlds. Hence, I had decided to build a one-off for this hackathon (which ofc can obviously be integrated with khaaliDimaag now that I think about it), and try and win the prize. The addition of claude code only accelerated that effort!

So, what was actually built? The project was a simple splitwise-esque app, but for crypto; ie, balances could be settled via crypto. Why? Well, many times I have had to send funds to friends over Coinbase to settle up balances. But also, think of the varied global conferences; oftentimes the attendees are coming from different countries, and don't always have access to the local financial infrastructure. With this app, strangers could pay for you and you can simply settle via crypto! Don't really need to touch fiat rails! & yes, I know p2p.me exists and is actually really cool, but I just wanted an app that is mine, ya know..

What else? One other thing I realized is that this can actually be a way to onboard folks into web3 as well, mainly because of Circle's Gateway Wallet. All my views on stablecoins aside, this is kinda useful. Add in some elements of yield farming or general DeFi, and you perhaps have a killer product! Also, how can I not mention ENS? This also becomes a way to then accept settlements from anyone else, anyhow, while not being tied to my app or platform (just my protocol / network)! True portable identity! Actually, this may solve an open problem with a different project of mine within the kdio ecosystem, since all user ids are effectively subnames of the TLD, so maybe there is a way to bake permissions in. With their NameWrapper shit, ENS has definitely been cooking lol.

Okay, enough yapping lol, here is the link for the live demo ~ khaalisplit.xyz (though how broken it is depends on when you are reading this blogpost). About claude code tho, almost entire code is written by it. Check out the github for commits by it (it really loves self-promotion lol, but fair marketing / advertising strat by Anthropic). The biggest thing is, it fucking wrote 300+ tests for the contracts! I am ashamed to say, I have not written as many in my ~15 years of being a dev. So, it is kinda clear why claude code is going to be a staple moving forward. At least for me.

Alright, I think this is enough for this week; there is some more stuff, but probably better to talk about them when they are closer to shipping state. In conclusion tho, claude-code, god fucking damnit (and yes, i am taking the lord's name in vain xD), this is a game changer! I have been seeing shit on X, and I have always been like "eh, but at their core, they're still LLMs and probabilistic", but god damn they can do some useful things. I had seen this tweet by @karpathy and kind of understood what he meant; but now I feel exactly what he means.. with a small caveat, I don't feel too "behind" (assuming his tweet is slightly pessimistic), but I feel much more excited! Because execution has always been my issue, and if I can leverage the tooling just right, oh boy, the things I could acheive! So, yes, excited is a more apt word for me; but I do agree with him on the "meta" layer or a new level of abstraction. More soon!