How I Learned To Stop Worrying and [sort of] Love the LLM

//01

Why is the adoption and integration of these tools such a contentious issue?
I think I've found the answer after finally signing up for my own Claude code account. 

Like I mentioned in my first post, I was pretty late coming to the LLM party, given how much of a computer person I have always been. You'd think I'd have been an early adopter. In man ways, I was.

I -- and nearly every professional creative I know whose primary environment is Adobe's Creative Cloud -- have been using generative AI in our day-to-day workflows for years. I can fairly confidently speak for nearly every professional designer when I say that the Object Select tool genuinely provides returns approaching the 10x or 20x degree. The ability to simply drag a box around a subject and create a non-destructive isolation mask saves us from some of the most (historically) tedious and frustrating parts of our jobs. Being able to expand a sky or a scene and adjust the composition of an image within it's frame -- other AI-based tooling I use frequently -- real, serious time-savers that enabled me to produce more work, better and faster.

Notwithstanding the ethical questions that remain unanswered around generative AI, no creative I've ever worked with is hostile to this kind of AI tooling.

Why, then, did I feel so weird signing up for a Claude account?
Why, then, did my stomach churn as I mashed out my prompt and I fed it my résumé?
Why, then, did it drop when it worked exactly the way I asked it to?

Was I hoping it wouldn't work?
Did I think that it shouldn't work?
Was it because I was wrong?

I spent the next two days working with Claude Opus 5. Prompt engineering came naturally to my overly-verbose and formal writing style and pretty large vocabulary.

I now not only have a résumé that is structured correctly for ATS systems, but is parsible across *multiple* systems and methodologies.
For the first time I can think of, I was able to design my résumé in InDesign and publish it as a .pdf file and it's being parsed correctly and consistently by every job board I feed it to.

Claude genuinely helped me diagnose, remediate, and address some *very* low-level .pdf issues that was going on with my document. Things Acrobat wasn't bubbling up at *all*, like text stream issues and how text was interleaving across different rendering engines. I was shocked and a little giddy as I watched it diagnose verifiable issues in the document visually, structurally, and from a content perspective and feed me actionable steps to address them.

It also lied to me three times, and did so while I had it set to "Max". Half a day, two Pro-level sessions worth of tokens, and 3 export cycles on my end were spent chasing a ghost bug in how it's tooling was parsing PDF artifacts.

"You're right." it told me. "That's on me...I owe you a correction going back several rounds."

I chuckle briefly at the accepted absurdity of a service that is capable of spending itself.

It was a delightful experience in this regard: I got the very best and the very worst in the same conversation and had the truth of its nature [relative to my own experience in a given domain] laid bare to boot. I had to tease out of the model clues pertaining to its own assumptions when generating responses from my inputs, and the only way I (or anyone else) can do that is by exercising our own domain knowledge.

Nevertheless, after weeks of sending hundreds of applications and feeling not only helpless but lost as to how to even address it, I finally felt like I was making real progress, at least. The document was tedious as all get-out to prepare whenever I made changes to the content but once set up and prepared, I had something pretty wild: a modest 2-column résumé .pdf that I'd designed myself which was being parsed reliably across Linkedin, Indeed, ZipRecruiter, Monster, and a whole host of other smaller board platforms almost entirely without error... even [and perhaps especially] on Workday instances. I've gotten more calls since my work with Claude on my résumé than I had in the months trying to use what had been pretty reliable years earlier. I'll distill the process in an upcoming post, because I know creatives like myself are really struggling just to hear back.

//02

While I started sending out applications with my new résumés, I began addressing what I knew was the other weak link in my professional profile: my website. I had spun up a minimum-viable-product following the layoff to support my job search, but I wasn't happy with the result. It felt dated and I'd clearly built it under pressure. Invigorated by the sense of progress and bolstered by the sudden uptick in response rate, I got to work... and immediately began running up against the most obnoxious part of the process: dealing with hosting providers.

I've used services like Flywheel in the past and they've been reliable, easy to use, and interface well with tools like Local that make setting up development environments for WordPress sites and staging environments really easy. They charge a premium for that convenience, though. When you're trying to save money like I was, things get a lot more greasy.

I won't name names but when the base install of a fresh site is pre-loaded with way too much plugin gunk and doesn't ever seem to work with the very backup plugin the managed hosting provider is pre-installing on my site, I can't help but wonder what else isn't working. Support chats tend to take five times as long as dealing with an issue internally. I was feeling pretty damn done.

I knew VPS options were orders of magnitude cheaper, but I just did not have the time to get fully up to speed on configuring the docker stacks from scratch I'd need to run something on a public server. I did, however, realize that I knew how to ask for it having stayed abreast of AI/LLM progress in the development space observationally and what's more: had played with enough of the underlying tech over the years and am proficient at consuming, understanding, and applying documentation myself to be able to check it's work, even if the line was a new one to me in the moment.

Did it give me something that worked out of the box? No. There were some hiccups along the way, a few rules that were misconfigured, but I went through that compose file and checked the commands that I didn't know or recognized were operating on sensitive aspects of the project. Most importantly, though: I took extra care to double check what it was telling me on anything security related. I even watched first-hand as a series of GET requests tried (and failed) to find compromised plugin files to exploit. The stack was preconfigured to deal with those outright without those requests even making it to WordPress. Nice.

I'll give it credit: it did in an afternoon what would've taken me, yes, perhaps 10 days or so to figure out on my own, and eliminated a lot of headache throughout. It effectively removed painful rounds of trial and error and gave me a well-marked path to follow. The trick is that I'm curious about these things and find enjoyment in learning them myself. I had no qualms about reviewing what it gave me because I could at least speak the language, even if I never learned the handshake.

I didn't use Claude for any of the design or hands-on site code work, though. This was deliberate. My website is in many respects a portfolio piece unto itself. If I used AI to generate the design or the code, what am I demonstrating? The question isn't about efficiency, it's about branding. I need to demonstrate competency in my crafts, and my personal ability to design (not to mention my personal command of language) directly and critically informs my effectiveness with LLM tools. I seek to demonstrate one with the site itself, and the other through my writing here: rambling, of dubious actual value to anyone but me, but nevertheless mine.

//03

I'm continuing to explore what I feel are the best use-cases for the technology to support doing the things I actually like to do myself. I don't think I'm going to be relying on Claude Design anytime soon but I'm going to explore it as a sort of bootstrapping tool. I've got a few things I'd like to spin up and this might finally be the way I ca do that and feel comfortable doing so while also being able to control my cloud spend and maintain the ability to locally-host whatever it is myself -- a home lab is most definitely in my future.

As for what I'm doing this week:  I'm trying to build a prototype MarTech stack that's not AI-forward, but Human-Forward with AI targeted only when and where appropriate, ensuring the actual cost to run the service is insulated against future rising token and subscription costs while still taking advantage of the remarkable (but nevertheless flawed) results this technology enables if wielded correctly and deliberately so. When I've got something worth sharing I'll share it; but whether it's going to be next week is a good question. 

I continue, of course, to seek out my next opportunity during this difficult job market, staying positive the only way I know how: with productive projects. This is where I must engage in the personal plug (which, I promise I'll stop doing once I'm on a W2 again): Hiring managers and recruiters looking for someone that can juice up your marketing and creative departments with some automation of both the deterministic and probabilistic persuasions,, please don't hesitate to reach out to me on LinkedIn as I'd love to have a conversation!