AI-assisted code generation and agentic implementation

How a little structure turned my chaotic AI coding experiments into a predictable, reusable workflow (and functional code).

I have to admit that when I first started experimenting with AI-assisted coding, I wasn't really sure how this whole thing was supposed to work.
I'd open a chat, throw a vague idea or question into the void, and hope something useful came back. Sometimes it did --most times, it didn't.

Surprise, surprise: this is not the best way to write code with LLMs.

Just seeing what sticks might be a brilliant approach for brainstorming, but, spoiler alert, it's a terrible way to run a coding project.

Once I learned how to actually set everything up, though, things changed quickly. The output became more predictable. The conversations made sense. It even started to feel useful? The problem wasn't the AI (although, yes, the models keep getting smarter every week). It was me -- more precisely, my lack of structure.

When I finally began to treat the model like any other programming tool, not like a magical search engine slash idea oracle, everything clicked into place.

I've been excited about a lot of things in my life before, but this time it's on another level.

I actively wrote code for about 15 years across multiple platforms. I used to LOVE writing code. I would skip meals or social events just to write. I'd even sleep less for it.

AI completely flipped that for me. Now, even thinking about having to write everything myself makes me anxious (seriously). I sometimes get short, almost panic-like moments just from the thought of it.

At the same time, I'm more productive and clear-minded than I've ever been. I recently quit my traditional job to make the biggest jump of my career, of my life.

People who don't use AI daily, or who still insist on writing everything themselves, are going to get left behind, too far behind (arent they already?).

Writing code, stuffing our brains with syntax, libraries, and APIs-- it was counter-productive. An unnecessary burden we carried for years. I'm grateful we don't have to work like that anymore.

Who can relate?

I finally did it.

I moved my blog off of Wordpress. It's running on a Go server on a small Digital Ocean droplet.

Why now? Because side projects are fun again. I'm excited about software engineering for the first time in about 15 years. Agentic coding is so new and unsolved! And even better, now I get to make political statements just by saying which agent I use. What a time to be alive.

Oh sure, I've had a lot of fun coding in that time. I've gotten excited about a lot of problems. But it was never software engineering itself. And side projects eventually burned me out. Especially once my toddler was born. I don't want to pick up a project, fight exhaustion, hit a weird error, and yak shave for an hour while hoping that I have fun tomorrow at least.

But wow, Claude[[0]](#footnotes) really fixed that. I'm not new to Claude or Agents; I've used some version of Copilot/Cursor/Claude at work since Copilot first came out, and I've been using agents for about a year at work. But work's different than side projects. At work, I can't vomit out 30,000 lines of code and hold it up and ask, "Is this anything?" But I now regularly do this at home as part of exploring how far we can push the tools.

And you know what? It's amazing for dad side project time. It can meet me wherever I am.

Is my wife walking to the store with our daughter? That's 20 minutes, I can write a prompt and let it churn once they come back.

Am I exhausted after both my daughter and my pager wake me up? I can just click through my project and whine about the parts I don't like, and Claude will dutifully fix it all. Or maybe I can just vibecode a huge project with the goal of learning something.

Do I have a few hours? Great, let's really bend this codebase to my will. I'm going to micromanage this to within an inch of its life.

And one of the famous slogans from "The Mythical Man Month" was "Build one to throw away." I.E. you should invest time to prototype before building the final system. Claude really changes the cost dynamics; you can build a prototype, prototype a second approach, prototype a third approach, refine the third prototype, and then the production system is within a stone's throw.

The actual Wordpress port

I've wanted to move onto a VPS ever since the [Wordpress drama](https://www.reddit.com/r/Wordpress/comments/1g2jo1j/explain_me_the_wordpress_drama_like_im_5/) happened years ago. But the juice never seemed worth the squeeze. I mean, I had fewer than 30 blog posts on this blog and just a couple of pages. Why bother, right?

But I first signed up for the Claude Pro account, and I tried thinking about projects that I might be able to one-shot within its narrow token budget. The blog port was a natural fit.

So I wondered if I should just convert the posts to Markdown and host them on [Github pages](https://docs.github.com/en/pages) or similar. But I liked the idea of being able to have dynamic server-based content[[1]](#footnotes).

Overall, I tried to one-shot the port at least 15 times.

In the beginning, I gave it really simple prompts. Basically, "Port [www.bitlog.com](https://www.bitlog.com)" to a Golang server with Markdown files storing content." These failed horribly! They'd just make a basic Go server and a few fake posts.

Next, I prodded it to download the content. It would try for a while but I would eventually run out of context. I tried asking it to make a tool to scrape each page, but it tapped out and asked me to export the XML instead.

So I downloaded the XML dump and started telling it, "The XML dump of a Wordpress install is in this directory." And my prompt grew and grew with each telling. So many things needed to be fixed. It linked to images on my remote server instead of hosting them. Pages included Wordpress styling. Opus 4.6's first attempt rendered completely blank pages.

At some point, I started experimenting with subagents and immediately started running out of tokens. This was the point where I upgraded to a Max subscription. That's how they get you and it worked. Well played Anthropic.

I then was looking for a [Beads](https://github.com/steveyegge/beads) alternative and found [beans](https://github.com/hmans/beans). I liked the idea of beads. I just wanted an implementation that... evolved a bit slower. Beans was another increase in power. My current experiments involve subagent teams, which are producing mixed results.

But eventually, I wrote [this prompt](https://gist.github.com/jakevoytko/ca2cb77f521b9f89e9c064596f7054c7), and I looked over the output. I realized, "This version has a lot of problems, but this is close enough. I can productionize this."

Refining the output

I started comparing the local markup with my Wordpress server. It skipped a bunch of meta tags like [OG](https://ogp.me/) tags, Twitter markup tags. I made a lot of changes to the visual design (graphic design is my passion), information architecture, etc. This kind of work was great when I was exhausted; I could just whine to it about not liking how the header was styled and it would go and fix it for me.

Then I started asking it to e.g. find accessibility issues. And it came up with some good ones, and suggested good things like having a "skip to content" element. And I noticed something funny! Whenever I commanded it to generate a list of issues that included severity, it would generate a list with 1-3 severe issues, 3ish medium issues, and 3ish low-severity issues. I find that I had to specify what I mean by "severe" for it to generate an honest list; like, "Judge all issues relative to a "severe" flaw that would render the site completely inoperable, like a focus trap."

Deploying

I created a Digital Ocean droplet, pointed DNS at it, and set up SSH keys so that SSH commands would work without needing in-band authentication. And then I told Claude that I wanted to set up Ansible and a reverse proxy, harden the server, etc. It churned for about 15 minutes, and at the end of it my blog post was deployed and all the configs were right.

And then I had to talk Claude off the ledge. Something about its environment was preventing it from seeing the page on HTTPS. I could access it just fine over both HTTP and HTTPS. And then we were live!

Conclusion

I converted a blog from Wordpress to my own Golang server, even though I don't have much time. I am excited about software engineering itself for the first time in 15 years. I have my own theories about how the profession will evolve over the next 5 years[[2]](#footnotes), which will be the subject of my next blog post.

First, obviously this took longer than doing it myself, given the number of iterations it took. However, Claude could work when I couldn't. It's indefatigable! On nights that I was too tired to code, but didn't want to play a video game, I could just whine to Claude and it would fix the problems I saw.

It was also a playground for a while. Almost a "code kata," except I wasn't trying to execute a perfect form. I was just walking down a well-worn path, seeing what happened each time I changed a variable or three.

But I did it. I deployed it. I'm happy with the results. And now I'm curious how far I can take this. Can I host my own email server?

Footnotes

[0] You can substitute your favorite agent here.

[1] To be clear, I never will have dynamic server-based content. But man, I love the idea.

[2] TL;DR: We need to become product managers before product managers become coders.

As a software engineer, I recognize that I've worked in a privileged field for a long time. The combination of good pay, flexible working conditions, and work that could sometimes provide decently entertaining challenges has been unique and frankly pretty awesome.

There have always been threats to the field, like outsourcing, H1Bs, and the usual capitalist megalomania that infects everything. But as a whole, there's a reason why a lot of people don't leave the field unless they're super burned out.

But now, with AI tool use being forced everywhere, everything just sucks. I fucking hate waking up and starting a job that used to be pretty good. Reviewing mountains of AI-generated code every day and never creating anything feels soulless and miserable. It's like being a doctor or therapist, but instead of seeing patients, you just review and compile their after-visit notes for insurance. You know, the part that sucks and everyone wishes they wouldn't have to do.

There are engineers who say things like "Oh this just lets me build things faster," "the code was never the roadblock," etc. but I don't agree. Regardless of where the field might end up going in terms of job totals, compensation changes, more offshoring, etc., the fact is that this technology came out and, for day-to-day engineers who care at all, it has just made being a software engineer fucking suck.

I don't know where I'm going with this, but I'm just tired of it.

AI vs Developer

I'm looking back from February 2026, and honestly? The headlines screaming"AI is killing developer jobs" feel overblown. I've lived the transition firsthand, and my career is stronger than ever.

I started as a developer in 2022, right before the AI boom. Landed in a fast-moving startup, working with Java (not exactly beginner-friendly). No senior dev shadowing me, no endless Slack help, my boss was slammed, so I couldn't ping him with every little question. That meant grinding through problems solo: debugging, reading docs, trial-and-error. Brutal at times, but it built unbreakable fundamentals. No crutches, just deep understanding.

Then ChatGPT landed in late 2022early 2023. Game-changer. Suddenly, 20% of my daily work  writing functions, handling tricky logic, boilerplate -- was AI-assisted. I was on free tiers of ChatGPT and Claude (premium felt like a luxury in startup life), but even that sped things up massively. In an environment where ship fast was the mantra, AI let me deliver quicker without sacrificing quality.

Back then, I used AI mostly for greenfield code. No GitHub Copilot access yet, so I didn't lean on it to explain or refactor existing team code. That kept me honest -- I still read and understood everything manually.

Fast-forward: I joined a multinational company. Boom  GitHub Copilot in the IDE, AI agents in the CLI, the full modern toolkit. My workflow evolved overnight.

Now it's simple (and powerful):

I drop the user story into an AI agent: Here's the requirement -- give me a thoughtful implementation plan.

It returns architecture, steps, edge cases.

I review, tweak for our system's realities.

Quick check with seniors: Does this align?

If yes-- I hit enter.

Features that took days? Scaffolded, coded, and ready to polish in minutes. Its not magic  its acceleration.

So, is my job gone in 2026?

If I blindly accepted AI output, merged without review, and called it done -- yeah, I'd be at risk. That path leads to garbage: misunderstood context, hidden bugs, exploding tech debt. We've all seen it.

But any developer with real system knowledge (even 50%+) wont do that. I dont. I treat AI like a turbo-charged junior: it drafts fast, suggests boldly. Then I bring my fundamentals -- architecture judgment, domain insight, trade-offs, security, scalability  to validate and refine. AI proposes; I own the decision. AI generates; I ensure it fits.

In 2026, thats the winning combo. Developers who master this partnership? Were not fading  were thriving. AI handles the repetitive 80%; we own the high-value 20%: critical thinking, system design, creativity.

My gut says were safe for at least the next 5 years. Beyond that? Tech always disrupts, and we adapt like we did with cloud, mobile, every wave before. New tools create new roles.

If youre-reading this as a developer: double down on fundamentals. Use AI aggressively as your co-pilot, never as a replacement. The ones who do will come out way ahead.

Views are my own and dont represent any past or current employer.

What's your 2026 reality? Has AI flipped your workflow like this? Juniors starting now vs. pre-AI vets -- what's different for you?

Drop thoughts/questions below
I'd love to chat.

Code review has always been my least favorite part of being a software engineer. Ever since we’ve started using AI at work though, I’ve noticed that most of my day has become reviewing code.

I genuinely don’t understand how some people are enjoying this more than coding by hand. Sure, debugging has gotten WAY easier but building things is just not as fun anymore. It’s like the difference between doing a puzzle yourself vs telling someone to do it and checking their work.

My theory: maybe I’m stuck in a loop of reviewing and correcting because my prompts are not precise enough. Maybe if I spent a lot more upfront time thinking about design and tradeoffs, that’ll get my creative juices flowing again.

Has anyone managed to get that “craftsmanship” feeling back in the age of AI?

Would you fly in a vibe coded airplane?

That question has been rattling around my head as I've watched the AI coding discourse split into two camps: "AI writes all my code now" and "AI is overhyped garbage."

The reality is messier. After months of building real projects with AI agents on GitHub Copilot, I've found that the developers getting the most out of AI aren't the ones who went all-in on autonomy. They're the ones who learned when to use it and when not to.

I wrote up the framework I use — four modes of working with AI, from hand saw to chainsaw. The key insight: staying in one mode all the time will slow you down.

A single line of code is faster to change by hand than to describe to an agent. Some things are harder to explain than to code. And sometimes you don't know what you want until you start writing it yourself.

The inverse is equally true. Hand-typing boilerplate when an agent can generate it in seconds is a waste.

The skill isn't using AI. It's knowing when to use which mode.

I've been using AI coding tools heavily for the past several months. This video is my honest take on what over-relying on AI has cost me as a developer and how I'm rethinking the way I work with it.

I've been using Claude Code daily for 6+ months. It's replaced most of my routine work:
• To understand code base - yes true this is game changing for me. its creates nice tech documents.
• Boilerplate generation
• Refactoring large codebases
• Debugging edge cases
• Writing tests & automation scripts - top notch in UT's
It outperforms Cursor and GitHub Copilot for deep reasoning and architectural changes. Period.
⚠️ BUT — always validate. Review the code. Run your tests. You're still the final gatekeeper. AI hallucinates less here, but "less" ≠ "never."

I work at a small B2B SaaS startup where leadership is getting "pro-AI" due to customer expectations. I'm not anti-AI (I use it daily), but our non-technical PM installed Cursor + Claude, built a quick demo app, and now wants to "prototype" features for engineering and even push vague initiatives like "use AI for automated E2E testing" despite us lacking solid testing strategy, acceptance criteria discipline, and clear system boundaries. I'm looking for advice on how to prevent reckless AI-driven development expectations from taking over before it becomes politically impossible to stop.