A podcast got me here. Episode 405 of Marketing Against the Grain. I was listening in the car, half-paying attention, and then I wasn't. By the time I pulled into the driveway I had already decided I was going to try Claude Code.
I had never opened a terminal before in my life.
The Setup
The first thing I did was buy the lowest subscription tier for Claude. Then I followed the podcast's advice and downloaded Ghostty, a terminal app, because apparently that's where this all lives. I installed Claude Code. I stared at the blinking cursor.
The terminal is an unforgiving place for someone new to it. No spell check. No autocorrect. No “did you mean...?” Just a cold error if you fat-finger a command. I'm still doing that.
What I Decided to Build
I knew enough to know I shouldn't start too big. I wanted something I could actually finish, something that would let me understand the workflow more than the end product. I chose a personal landing page for a project I've been thinking about, hosted at cliqology.com.
But the real goal wasn't the website. It was understanding how the pieces connect: Claude Code writing and editing files on my MacBook, me pushing those changes to GitHub (a place where code lives in the cloud with a full version history), and then GitHub automatically sending the live site to Netlify (the service that actually serves the page to the internet).
I wanted a safety net before anything went live. That's why I chose this workflow. I wanted to stage changes, review them, and only push when I was ready. That instinct turned out to be right, though it added some complexity upfront.
Where It Got Expensive
Here's what I didn't anticipate: I burned through 70% of my monthly token allocation in the first six hours.
Tokens are the unit of work Claude operates in — roughly, every word you send and every word Claude responds with costs tokens. I was sending long, conversational prompts. I was going back and forth. I was re-explaining context Claude already had. I wasn't being efficient, because I didn't know what efficient looked like yet.
Now I'm using Claude.ai to help me draft my prompts before I send them to Claude Code. Yes, that means using one Claude to talk to another Claude. It sounds absurd. It works. The idea is that a well-formed, specific prompt uses fewer tokens and gets better results than a rambling one. I'm still early in figuring out what that actually looks like in practice, but the early intuition feels right.
The Flow State Problem
There's a thing that happens when this starts working. Hours disappear.
I sat down to work on this and looked up and it was dark outside. I hadn't noticed. There's something about having a capable collaborator available at every moment that removes every natural stopping point. You fix one thing, you see the next thing, you ask about that. It keeps moving. That's exciting and also a little dangerous when you're on a token budget.
What I'd Do Differently
Start with a tighter scope on the prompt. My early prompts were basically conversations — open-ended, exploratory. That's fine for learning, but it's expensive. I'd have spent more time upfront writing a single, clear prompt that described exactly what I wanted before sending anything.
I'd also have budgeted my tokens consciously from day one, the way you'd budget any other resource. I didn't think of it as a finite thing until it was mostly gone.
Where Things Stand
The site is live. The workflow works. I'm learning how to prompt more intentionally.
I still type wrong commands into Ghostty more often than I'd like to admit. The terminal still doesn't care.
If you're thinking about trying this: pick one small thing and build it. Not a big thing, not an ambitious thing — one thing you could plausibly finish in a few hours. Focus on understanding the process, not impressing yourself with the result. The result will get better. The process is what you're actually learning.