How to Be a Wizard: AI-Assisted Coding and the Art of Leverage

I always play a wizard.

Back in the 1990s I loved Diablo. I played it through as a Warrior, a Rogue and a Sorcerer. After trying out each path (and getting some not-so-legit items from others on BattleNet), I realized wizards were the way to go. I could clear an entire room in seconds – or die spectacularly fast.

Using a Staff of Apocalypse, I’d hit all enemies on the screen at once. Later on, I played through Ragnarok Online as a Wizard. Then Skyrim. If I wasn’t a wizard, I was a rogue – switching MP for arrows but keeping a distance from my enemies.

I’ve always been drawn to glass cannons. Wizards are optimized for dealing damage. They’re frail. They can be taken down easily. But when used right? They’re a powerhouse.

I’ve kind of applied the same mindset to my career.

I optimize for being as productive as possible — using whatever tools enable this. That philosophy led me through PHP, ColdFusion, jQuery, WordPress, Ruby, and for the last 20 years, Ruby on Rails. Each was the most powerful spell available at the time that I knew of.

More recently, the biggest upgrade to my damage output has been AI-assisted coding with Claude Code.

Leveling Up: The Spellbook Gets Thicker

Most of the senior developers I’ve worked with spend time polishing their tools. They setup keyboards, custom shortcuts, and tweak their code editor in a way where they can maximize their focused time.

This ranges from the tools we choose (languages and frameworks), to the software we write in (code editors, docker), and more recently to how we leverage AI to write faster (LLM models, agents).

Each level of abstraction take a step back from tweaking bits and bytes. Improved languages and frameworks led to a major increase in what we could do in single coding session.

That’s a measurement I’ve never thought about, but it becomes more clear the more you can do in less time.

Writing jQuery was loads faster than writing all of that pure JS DOM code.

Creating a basic application in Rails, complete with it’s conventions, allowed for setting up an app with fewer decisions and get to creation that much faster.

And now with Claude Code, languages and frameworks are mostly abstractions. I am still in the “mostly look over the code it writes” phase of AI coding, but I suspect it’s only a matter of time before that domino falls and I’m shipping (more) code to production without human eyes.

Each of these jumps in productivity felt like a level up. Like I’d invested in a new skill path, unlocked a new ability and everything was smoother.

Recent Dungeons

The last few months, since around the beginning of December 2025 when Anthropic released the Opus 4.5 model, have felt different. It’s no longer a question to me if I’ll be more productive with LLM-assisted coding. It’s how much more productive I’ll be.

My LLM experience isn’t too different from most developers. I started by asking ChatGPT some questions in 2023ish. By 2024 I’d tried out Cursor a bit. By July 2025 I was using Claude Code. Over a weekend I built out a prototype of the Community features on Hardcover that I’d been dreaming of. The code wasn’t the greatest, and neither were my systems at the time. It was a throwaway branch, but it showed me what was possible.

Fast forward to December 2025 and the tools, along with my ability to work with them, have improved. I’ve launched a bunch of things in only a few short months.

Often what would take a week of work can now be done in an evening.

Each of these projects would have taken longer without AI assistance. Some might not have happened at all — the activation energy would have been too high. The change to this blog (from static to dynamic) had been on my mind to do for MONTHS, and I wrapped it up in an afternoon while watching Netflix.

That ability to finally execute on those open loops in my head have been a welcomed relief. Sure, I would’ve been able to do any of the above. But they would have taken significantly more time, and involved much more frustration.

In the time it would previously take to research how to do something, LLMs can propose and implement a solution.

Glass Cannon Problems

But here’s the thing about wizards: they’re fragile.

Optimizing for maximum damage output means accepting real vulnerabilities. The same is true for leaning hard into AI-assisted coding:

If you can’t write a for-loop without AI anymore, that’s a problem. High-level wizards sometimes forget their cantrips. The basics still matter when the magic fails.

AI is confident. It writes code that looks correct. But hallucinated APIs, deprecated methods, subtle logic bugs — the illusion of correctness is dangerous. You’re responsible for code you ship, even if you didn’t write every line.

Granting AI access to your systems is summoning power you don’t fully control. A misconfigured agent could delete a production database. Could leak customer data. Your creations can turn against you.

Wizards without mana are just squishy people with sticks. If Claude goes down, Anthropic rate-limits you, or your API budget runs dry mid-sprint — suddenly you’re coding at half speed. What’s your fallback?

The classic trope: power changes you. I’ve noticed I spend more time preparing to code and reviewing code than actually writing it. There’s a risk of losing the joy of crafting code yourself. Of becoming a code reviewer instead of a coder.

AoE spells hit your own party. AI “helping” by refactoring things you didn’t ask it to touch. Cascading changes. The blast radius of a powerful tool wielded imprecisely.

Solo Leveling Your AI Skills

In Solo Leveling (the anime), Sung Jin-Woo doesn’t start as the strongest hunter. He’s E-rank. Weak. But he gets access to a system — a way to grind, to level up, to become something more.

Learning to use AI tools effectively is the same grind.

You don’t install Claude Code and immediately ship 10x faster. You:
– Learn what it’s good at (and what it hallucinates)
– Build intuition for prompt crafting
– Develop review habits that catch bugs
– Figure out when to use it vs. when to just write the code yourself

Every project is a dungeon. Every bug is XP. There’s no shortcut to S-rank.

One of my favorite fun fantasy books is Off to Be the Wizard. It’s about normal guy who figures out how to tweak reality; which somehow leads him back in time to the middle ages where he pretends to be a wizard.

Likewise, even having access to reality-warping “magic”, it takes practice and imagination to use it. We’re in the same place now with LLMs, where it’s unclear what the ceiling will be in what these systems can create.

The Tower Spirit

⚠️ SPOILERS for Rhythm of War (Stormlight Archive Book 4) ⚠️

If the wizard metaphor doesn’t resonate, here’s another frame for fantasy readers:

In Brandon Sanderson’s Stormlight Archive, there’s a tower called Urithiru — an ancient, magnificent structure that was dormant for millennia. What awakened it was the Sibling, a spren (a sentient spirit of Intelligence) that bonds with the tower and brings it to life.

The Sibling:

• Controls the tower’s systems — heating, lighting, defenses
• Lay dormant until someone came along who could bond with it
• Partners with Navani (a scholar/engineer) to experiment and create new fabrials
• Provides capabilities the tower couldn’t access alone
• Still requires a Bondsmith (human) to provide direction and Investiture

The parallel writes itself. The AI is the tower spirit. The infrastructure (Mac mini humming under my desk) is Urithiru. And I’m… Navani? Experimenting in the tower, trying to figure out what this thing can do, occasionally making breakthroughs that feel like magic.

The delightful absurdity: an ancient, powerful tower spirit… running on a small aluminum box in a home office.

The Disclaimer

LLMs and AI aren’t all roses. Most of these models are trained on stolen data (and are running on stolen land using stolen water). The data infused in these models is ill-begotten. They use too much electricity and water, displacing local communities. They take jobs from people by moving what a single person can do. They reproduce artists works, copying skills that took a lifetime of practice. Some companies are pandering to big tech and fascist governments.

And yet those same tools can compress 3 weeks of programming into a single day, already. Time will only make those tools more impactful.

Even as someone who aims to avoid big tech when I can, it’s hard to ignore the productivity gains available. I’m ethically split on the subject – allowing myself to use these tools while knowing it’ll be a while before these concerns are addressed (if ever).

The Spellbook Isn’t the Wizard

Anyone can buy a spellbook.

But the wizard is the one who spent time learning to read it. Who knows which spell to cast when. Who understands why the incantation works — so when it fails, they can adapt.

It’s one of the most exciting times ever to be a developer. What we can do is limited by our imaginations and time.

AI is the most powerful spellbook ever written. But having access to it doesn’t make you a wizard.

You make you a wizard.

The tools are available. The dungeons are waiting. The only question is: are you ready to try?

Related Writings

• Scenes from an Agentic Life by Thomas Dohmke
• Living in the Inflection Point by Brittany Ellich
• Do not give up your brain by Cassidy Williams
• Stop generating, start thinking by Sophie Koonin