POTPOURRI INTRO DEV
POTPOURRI
So what, if anything, can LLM "vibe coding" actually do RIGHT? Not a lot, but maybe a little...
I'm going to preface this article by stating this truth : I'm still 100% team "don't trust LLM vibe code for anything remotely mission-critical." And I promise your sweet ass I'm still light years away from firing up ClaudeAi/ChatGPT on the daily. But I did recently stumble across at least one use case that saves me a notable amount of time and generally doesn't trigger my "coding burnout" sensors : low impact web coding.
Before you gather up your virtual pitchforks, let me be absolutely clear : if you use LLMs to displace workers/creatives, you totes suck. Also, and I can't stress this enough : vibe.. coding.. is.. not.. real... coding.. and.. current.. LLMs.. are.. not.. worth.. the.. environmental.. harm.. they.. cause. LLMs are truly bullshit machines, and if you rely on them for nuanced coding tasks (or any real information, for that matter) you get the dumb prizes you deserve for playing those sorts of games. We are hitting "much scarier than the .com ruin" territory with these suckers, so much so that even Altman, Zuck, and Bezos have said we are reaching "financial bubble range" (without, of course, admitting the faults of the models themselves). This is undoubtedly going to end miserably for a lot of folks and entire cities for that matter. That said, just because LLMs aren't as ground-breaking as these techboi carnival-barking edgelords swear they are, does that mean they are complete ass?
We know at least one thing about them : as with most ML based models, LLMs tend to perform best when you feed them with earth-scorching, metric ass-tons of accurate data. And every LLM that scrapes webpages for their training (which is, to say, every major LLM and even ones you haven't heard of if my website datalogs full of thirsty ass model headers are being truthful) has an over abundance of 3 things that are generally pretty accurate and well-formed because ~70% of the sites on the internet use some form of CMS:
HTML, CSS, and JavaScript.
And we haven't even touched on the fact that many of the most trafficked sites on the internet are bespoke, professionally developed non-CMS sites as well. In other words - most of the HTML, CSS, and JavaScript ingested by LLMs are pretty damn good.
This hits an odd sweet spot for me for several reasons. The coding I enjoy doing is generally "large scale data" stuff as well as game mechanic design. Stuff I hate to fudge with and honestly aren't very good at : UI and UX ("no shit!", you probably say, as you read this blog post..). Somehow, I've maintained a small but surprisingly durable little community for a web-based game (27+ years and counting with a domain name registered before Google.com). The community I serve is still vibrant enough to allow us to get a small Twitch stream together with 30-50 players joining in weekly to listen to me and my game cohorts say dumb shit and give some game stuff away.
Initially, we used a "spin wheel" we found on the internet to assist with those giveaways, but the randomness of the resulting winners seemed suspect at best. So I had a(n admittedly nerdy) plan - why not create a wheel that's based on pseudo-random MD5 hashing to determine the winners? That way, if someone doesn't believe the "wheel results", the chosen-winner can be accurately and publicly audited. So I created a hilariously bad-looking MD5 wheel that did exactly that (that I will NOT link to...). It worked, but (again) it looked rotten.. but it WORKED.. and cleaning it up was secondary to accuracy.
But after getting (understandably) ribbed one too many times on the stream for its aesthetic, I wondered to myself : could I have an LLM help make this hot-but-working-mess of a spin-wheel presentable? Again to reiterate - I would NEVER allow vibe code to be placed directly into actual game code (glaring examples of even low-impact failures follow below that doubly confirmed this fact to me), but this seemed like a relatively low risk sample use case.
So I fired up ClaudeAi (Sonnet 4) and "described" the problem I faced. Surprisingly, it gave me a pretty good-looking result out the chute.. I could see why some people would be immediately smitten!
..but "looking good" and "being accurate" are 2 entirely different things...
- The celebration screen announced a winner that didn't match the wheel spin winner
- My lizard brain auditing proved the MD5 hash they displayed was being calculated incorrectly
So, while it looked much better than my shitty wheel, almost nothing it created really "worked." Fabulous.
I messed around a bit more and burned through my free tokens quickly. So I tried again later and ClaudeAi was just shitting the bed left and right and the wheel quit working altogether... it had even inserted some stray, misplaced JavaScript visible at the bottom of my page. This should have been an easy offramp for an LLM hater like myself to decide upon, but I instead picked the strangest option in this self-loathing "choose your own adventure" drama...
I took the code ClaudeAI created and sent it over to ChatGPT (4) and, shockingly, it cleaned up the MD5 formula and the page once again worked. I couldn't get it to "fix" the celebration page however (or the fact the "wheel" spun fine the first time but never completed a full rotation later), so I got to digging through the JavaScript code myself. I modified the wheel to switch directions between left and right on alternating spins which fixed the "non-spin" issue (it was because the angle going past 360 and a "spin count" variable not resetting). I also noticed by random chance the "wheel winner" was always 90 degrees away from the winner stated on the celebration screen. So I just coded things to tilt the wheel 90 degrees... and everything suddenly worked and looked great!
I used this version of the wheel for a week or two on the stream to positive results... but it was incomplete. So I decide to drag the working code back into ClaudeAi to make a small addition - an "end of game" jackpot award that I had been checking manually.
Aaaannddd.. it broke the MD5 formula again in the process. WTF? Fortunately, I got things working quickly this time using only ClaudeAi (though I did have to get a notably verbose about how it screwed things up in the process). Soooo...
Here's a link to the working MD5 Wheel Spinner that now does (or at least seems to cosplay) exactly what I wanted it to. The code inside the page is roughly 85% Claude/ 12% ChatGPT / 3% Me.
This all sounds like a huge pain-in-the-ass, but it did allow me to walk away with a much more nuanced and well-formed opinion on how LLMs can be used successfully for small web projects with low impact that need zero scaling. But, with positives, come a few caveats :
- if you aren't willing to audit shit, you're setting yourself up for a bad result
- unless you know how to code to "baby things along", a lot of shit will still break and almost always be imperfect at best
- it did a LOT of the work I didn't want to do/relearn (primarily "pretty" CSS UI code) and would have never bothered to do, and gave me an overall improved product
- considering I'd have had to do all of this myself had I decided to do it (big nope), it absolutely saved me 3x+ time
- my little indie user game base is happier
- This didn't burn me out mentally like coding from scratch at the moment
..and after this initial experience, I actually found a couple of other use cases for ClaudeAI / ChatGPT where they save me quick spurts of time in my REAL job -
- If I have the CREATE statement to a wide SQL table and need to create an initial insert statement (no, I'm not going to use INSERT *)
- If I have a field list in one language (like HTML, JavaScript, Python, or SQL) and need it converted quickly to the other
- It's good at creating Python lists and dictionaries from data
- If I need JSON field quickly plucked out of a formatted JSON string, it's solid
..and that was it.. until yesterday when one of my favorite bloggers / social media folks posted like his 17th blog posts since I posted my last, and motivated me to get my ass in gear.
So yesterday, as a refresher for this here blog post post, I tried the new Claude Sonnet 4.5 model to create the following quick webgames for my game :
Hash Scrapper - an alternative game for my Twitch players to participate in to choose an auditable winner using MD5 hashes in a modified "battle royal". Pretty sure this was 100% Claude Sonnet 4.5; huge time save; easily 5x+ savings
Steal or no Steal - a "post wheel" prize game I plan to sporadically use. This one was a little stickier but still had what I'd estimate to be a 3x+ time save. But to get there, things :
- Shit the bed a few times during generation.
- Originally gave away the winning box accidentally.
- Boxes were in order "high value to low" instead of numeric order by box number.
- Original shuffle I asked to implement wasn't suitable so I had to manually intervene
- I had to drag it between Claude 4.5 and ChatGPT 5 when Claude tried using a function name with the same name as a variable name repeatedly.
- The model lost part of the CSS code somewhere along the way and I had to manually create a few lines.
- I also had to modify JavaScript to reset a variable it forgot about.
- My previous experience babying the MD5 Wheel helped here
2025 Triple Crown Horse Selector - All ClaudeAI 4.5; it did the research of Triple Crown racers and created a working page within like 2 iterations. I know horse racing pedigrees relatively well, and it looks like it probably got the research right but honestly if it didn't it doesn't really matter for game giveaways... 6x+ savings
..so that's it! Per the usual : Use any and all of these webpages however you would like (but of course at your own risk)!
Conclusions -
- I would have only ever written MVP versions of this stuff manually that would have looked like shit without LLMs and I'd have been fine with that
- I'd never pay someone to do more than the MVP version I would have written manually
- I fed LLMs precise "logic" on how I wanted things work in almost all instances; I have a bad feeling they would have gone rogue had I left them to their own devices
- My player base (which supports my niche, indie ass game) is having a better time because I mildly use these tools
So no - I'm not going "full LLM honk", and I will absolutely openly shit on any company replacing workers and creatives (trust me, my blog art is my own 
) with LLMs. But I also drive less than 2000 miles a year (about 1/7th of an average US citizen) so I'm not going to feel overly bad about running a few LLM queries now and again to keep my loyal indie game supporting playerbase happier as well as keeping me from burning out in the process when every damn company is cramming LLMs into every orifice of every fuggin' thing we use anymore for no good reason.
) with LLMs. But I also drive less than 2000 miles a year (about 1/7th of an average US citizen) so I'm not going to feel overly bad about running a few LLM queries now and again to keep my loyal indie game supporting playerbase happier as well as keeping me from burning out in the process when every damn company is cramming LLMs into every orifice of every fuggin' thing we use anymore for no good reason.And yes - I'd happily give up my handful of LLM usage tomorrow if the whole damn thing disappeared into the ether and took the environmental and psychological damage they bring with them, but I fear that particular LLM horse has sadly already left the barn.. thanks again techbros and capitalists!
2025-10-06 11:13:21
