When I started using Claude Code, I wanted to see how the experts worked with it, the same way I used to read video game guidebooks on the way home from the mall as a kid.
I found Dex Horthy’s talk on No Vibes Allowed and checked out his repo. I copied his commands into my local setup and typed / to see them all. 27 new commands showed up, including four ways to research, eight ways to plan, and multiple implementation options. I kept asking: why is this here? And when do I need it?
I ripped everything out and went back to vanilla Claude Code.
Building my own workflow started with em-dashes. Every response came back riddled with them. I’d delete them, get the next response, and they’d be back. Then “quietly” started sneaking into everything. Bolded bullet-point lists where prose would’ve been fine. “It’s not X, it’s Y” constructions that no one would use. I got sick of fixing the same things, so I wrote them down as rules and that became a writing style skill.
One thing at a time
Years ago, I wanted to improve at StarCraft 2. Games lasted twenty minutes of locked-in concentration, and I’d have a headache after each one. But playing more wasn’t helping me win.
I went looking for holes in my gameplay, watching replays, and writing notes on what I could’ve done differently. It’s painful, like listening to a recording of yourself. Why did I initiate that fight? Where did I stop making workers?
One leak kept showing up: I forgot to spread creep when things got hectic. Creep is terrain that gives your army a speed advantage, but you have to expand it manually. I set a 30-second repeating timer. When it went off, I stopped whatever I was doing and checked on creep. I knew it was working when I caught myself checking creep without the timer going off.
That process got me to diamond. Then I worked on another leak.
One tool at a time
One of the first things I tried in Claude Code was Plan mode, which tells Claude not to write code yet. I described a feature I wanted to build, it asked me two or three questions, and spit out a thousand-word design doc. I thought, “I’m not reading that.”
I used Plan mode for a few features anyway. Each time, my brain turned off, and I’d force myself to read the whole thing and offer notes. Claude regenerated the plan, and I re-read everything to find what changed.
The Claude reddit kept recommending a plugin called superpowers. I installed it for its brainstorming skill. Instead of generating a huge plan, it asks you questions. “You’re building a URL shortener, cool. Is it just for you or other people?” Then: “Great, just for you, so we don’t have to think about scalability. What about the database?” After the questions, I could see each piece of the design, steer it, and it only combined everything at the end.
I didn’t know I needed that until Plan mode failed me.
Another leak
I used Claude to review my own code before submitting it. It caught things I’d been missing, so I ran it on other people’s code too. I fed their changes to Claude, read its comments, and used them in my feedback. A senior engineer responded to one of my comments: “Great catch, I would’ve missed that bug.” I hadn’t caught it, Claude had. But I didn’t say anything.
A “good note” on another review made it feel worse.
So, I built an agent skill for code reviews and made the bot start every review with a robot emoji 🤖 and a note that I would follow up with human review. As soon as it posted the bot review, it opened my browser so I could follow up with my own comments. I went back to that engineer and told them I’d used Claude on earlier reviews.
Over lunch
My brother-in-law Brian said, “Right before I left to meet you, I kicked off a design shootout with three front-end agents, then toggled between them to see what I liked. I wouldn’t have done that when I just started.”
I asked if that was something you could teach.
He said: “Legendary guitarists talk about how much Eddie Van Halen influenced their playing. But when they finally got a chance to play with his rig at a soundcheck, they sounded nothing like him. They only sounded like themselves.”
Yep.
Thanks to Brian Sandoval for feedback.