Wands at the Ready
On the fence
Paul was building a fence outside his Mountain View home. His son, Steve, was at his side. "You've got to make the back of the fence, that nobody will see, just as good looking as the front of the fence," said Paul. "Even though nobody will see it, you will know, and that will show that you're dedicated to making something perfect."
Fifty years later, Steve Jobs returned to his childhood home with Walter Isaacson, his biographer, and inspected the fence that left an imprint on his entire worldview.
I'm no Steve Jobs, but the lesson that Paul taught young Steve resonated with me. The instinct to paint the back of the fence even though nobody will see it, to design and build something both functional and beautiful, has always been part of my DNA.
Ideas have piled up over the years. A row of half-filled Moleskine and Leuchtturm1917 notebooks in my office can attest to that. 2016: "the Vault: a secure digital storage facility for retirement plan sponsors and administrators..." 2018: "mycaddy: an AI-powered app that serves as your caddy on the golf course. Voiced by famous caddies." (I didn't say they were all good ideas.) I knew what to build. I just couldn't build it.
Learning to code seemed daunting and time consuming. Hiring a developer required capital I didn't want to commit. The costs to build were clear: time, and money. Deciding which one to spend usually paralyzed me just long enough to let the idea fade into the background, replaced by a new, shinier one.
I've grown accustomed to life on the fence.
Wandering into the wand shop
Until recently, my experience with large language models (LLMs) was limited to chatbots. They were entertaining, but lacked real-world, business application.
Thankfully, someone in a lab wondered, "what happens if we give it tools?"
Boris Cherny, the creator of Claude Code (now Head of Claude Code) at Anthropic wondered 'what if...'. What would happen if you gave Claude a special toolbelt with a tailored operating manual aimed at helping developers be better developers? It turns out, a lot.
When Anthropic launched Claude Code to the public, word spread fast. First among the developer community, then more broadly. The YouTube algorithm, no doubt informed by my "how to build software using no-code tools please help and thank you" searches, went to work on me.
I didn't understand the technical side of how it all worked. But I saw clearly the potential it unlocks for anyone who learns how to use it. I had to try it.
Claude Code isn't for the faint of heart, at least at first glance. It's an application that you have to run via a 'terminal': that thing that all of the hackers in all of those movies use to hack the system.
I Googled "how to install macOS Terminal" only to discover it comes pre-installed. Beaming with confidence, I moved on to "how to install Claude Code."
The first sessions were tentative. I wasn't sure where my data was going, what this tool would have access to, or whether I was about to accidentally email my entire hard drive to the internet. Like walking into the wand shop on Diagon Alley for the first time — equal parts impressed, excited, and genuinely unsure whether the next thing I touched would blow up in my hand.
I clicked around. I made mistakes. I wasted time formatting my Terminal color scheme. But mostly I just...took it for a spin. I asked Claude to build things and got back impressive results, but I still had no idea what was going on under the hood.
I felt like Bambi skating around on the ice.
Learning to ask for directions
When you enter a new and strange land, one of the first things you need to do is learn to speak the language. You don't have to know how to write poetry. You do need to know how to ask for directions.
I bookmarked the official Claude Code documentation page, going back and forth between that site and my newly christened Terminal app for hours and hours on end. Slowly, my orientation took shape. I learned the difference between an agent (I'd heard of that) and a skill (never in this context). I learned what a "primitive" was and why it mattered.
The single best tip I can offer anyone starting out: remind Claude to consult its own documentation whenever there's a question about how to implement something. Claude Code has a built-in reference agent that knows the official docs cold called claude-code-guide. Telling your AI coding partner to check its own manual before guessing will save you hours and a startling number of tokens. I wish someone had told me that on day one. Once I figured that out, my first real project went from idea to production in a week.
I am so smart. S-M-R-T!
The eureka moment came when I realized you can, and should, use Claude Code to make Claude Code better.
This happens one of two ways:
- doing the thing helps Claude learn how to do the thing better
- you can modify the tooling and update the operating manual to customize how Claude does the thing. Every plugin I build makes the next one easier to build, because the plugins themselves improve the tool doing the building.
Fair warning: You get a surge of dopamine when you write software that helps the world's best coding AI learn how to write better software. It's slightly addicting.
To keep yourself grounded, remember the fate of Icarus.
Or Homer Simpson.
There's a scene from The Simpsons where Homer, after becoming "a college man," decides he doesn't need his framed high school diploma anymore so he lights it on fire. He dances a little jig and sings, "I am so smart. I am so smart. S-M-R-T! I mean S-M-A-R-T!" as his living room burns around him.
It's easy to confuse "tweaking" with "building." That's the biggest trap — for me at least — of working with something like Claude Code and its nearly limitless potential.
Now, thanks to the expert abilities of frontier models and clever, sophisticated harnesses like Claude Code, we are all S-M-R-T.
There are no more excuses. Coding is solved. You and I can now build software that solves the rest. And if out-of-the-box doesn't suit our needs, we can modify the toolbelt and operating manual to fit our needs.
What I actually built
Four plugins. smrt plugins. Each one solves a problem I ran into.
smrt gives Claude institutional memory. Without it, every new session starts from zero — your AI has no recollection of what it learned yesterday, what mistakes it made, or what patterns work in your specific environment. smrt captures those learnings automatically and compounds them over time. Claude gets sharper with every session because it's building on what came before.
smrt-powers handles engineering operations. A nod to 'superpowers', smrt-powers manages backlogs, tracks progress, runs session handoffs, and keeps the development workflow structured. When I'm building something new, smrt-powers makes sure Claude knows what was done last session, what's next, and what needs attention. It's the project manager I never hired.
smrt-vision is the operations layer. I run multiple projects and early-stage business entities — smrt-vision gives Claude awareness across all of them — strategy docs, action items, blockers, priorities. One command in the morning and I get a briefing that covers everything happening across my portfolio. It's the chief of staff role, automated.
smrt-audit handles security. Every piece of code Claude writes gets checked against security best practices. I don't have a security team. I don't have a security background. But I have a plugin that makes sure the worst offenders are covered before anything ships.
Together, these four plugins turn Claude Code from a capable coding assistant into something closer to an operating system for running a business. They talk to each other. They build on each other. And because they're built inside the tool they extend, they get better as the tool gets better.
It's tempting to say, "I built this without writing a single line of code," because that's what the hype-machine demands. I prefer a different view: Working together, Claude and I built these plugins to help the human+AI collaboration get a little better each day.
Full-stack engineering might be dead. Full-stack partnering is the future.
Wands at the ready
The hard part isn't the code anymore. It's knowing what to build, being able to describe it clearly, and giving the tools what they need to validate the work.
In a recent Enterprise Briefing, Kate Jensen, Head of Americas at Anthropic, referred to "the magic of Claude Code." It's natural to hear the word magic and close the browser tab. I get it. The AI hype cycle usually under-delivers.
But after spending months inside this tool — building, breaking, rebuilding — I'd struggle to find a better word for the moments when it clicks. When you describe a problem in plain English and watch working software appear. When the tool you built yesterday makes today's build faster. It does feel like magic.
Yes, Claude still makes mistakes. Yes, the constraints of the system — the primitives, the context windows, the occasional over-confidence — mean we're still limited. But the wand shop is open, and I for one am willing to give it a go.
Just...someone let me know if my living room is on fire.
Cheers, Michael