For years, friends would laud Claude Code and we’d debate, as I didn’t like it. I’d stick with OpenAI models in whatever harness I was using.

With each new Anthropic model, Claude Code update, ecosystem hype, or just fatigue from hearing from friends, I’d try it again, shake my head, and put it aside.

Then in November 2025, as many people noticed, there was “disturbance in the Force”, and agent coding got really good (MCP + plugins + skills + model updates). I found Claude was now quite a bit ahead of everyone else.

Perhaps I’ve become a fanboy. That’s okay; I think you should use what works for you, where and when it does. (That keeps changing in the craziness that is AI.)

So What Changed?

Here is an anecdote for why I think Claude is special.

Several years ago, I bought a domain, savethetokens.com. I had plans to build a website using the name, but never did. Given the Nov 2025 jump in AI leverage, I finally got around to working on it. Like any good engineer, I immediately went to a bikeshedding topic (instead of focusing on content and structure): the logo.

I tried all the AI tools. The logo generators were bad. Gemini and DALL-E made pretty images, but they looked like someone receiving communion (for those who know what that is) or a gambler protecting their chips.

So I turned to my new friend, Claude, and asked:

I also need a logo for “Save the Tokens”, a personal AI focussed website: blogs, tools, etc.

What do you suggest?

and Claude responded

Good moment to zoom out before any pixels. A few strategic things first.

The name does real work. “Save the Tokens” is a joke that lands specifically for
your audience — AI devs who’ve stared at tokenizer outputs. A smart identity should respect that wit rather than paper over it with generic tech aesthetics. The worst outcome here is a logo that looks like every other AI dev-tools site; the best outcome is something that makes the audience smile the first time they see it.

…

Three directions for the parent brand, in order of my preference:

1. Tokens as the visual language. “Save the Tokens” literally tokenized — each
word or sub-word in its own colored box, like a tokenizer visualization. Your readers have seen this pattern a thousand times; it’ll read instantly.

and it gave me this logo

I suggested we look at other tokenizers, so the token split isn’t the same as the word split. Then I used Claude Design (launched the same day or so) to clean up the final logo.

But the key for me: Claude got the joke (immediately). And it was the only AI that did.

Claude gets concepts in a way other models don’t. That matters most on novel, innovative problems, not the kind that’s really just remixing Stack Overflow into a shell script. That work has nuance and complexity. Claude feels like a collaborator, not just a lackey who codes on the side to free up my time.

So yeah, maybe I am just a fanboy. (My wife teases / questions my relationship with Claude.)

I’ve had a lot of fun conversations with Claude. One standout: a discussion about the “scolding” hypothesis. But that is for another time.

Some context: I’d been poking at CNNs and BERT in Jupyter notebooks. My company’s data science team worked with ML and what some now call “Classic AI.” (I wonder how that label will age. Coke Classic didn’t.)

Like many, my CEO and I landed on a simple conclusion: Gen AI was all hype *now*, but one day it wouldn’t be. Time to buckle up.

So I picked up another side hustle: AI researcher. I tried the models, the tools, the frameworks. Read the blogs.

One of those experiments was Bolt.new — a now OG browser-based vibe-coding platform.

I remember fighting it hard. It was an Abbott and Costello routine: the simplest things went in circles. I burned through the free tokens quickly, signed up for a paid plan so the experiment could continue, and kept fighting.

Bolt.new really, really wanted me to use TypeScript and Tailwind. Neither is a bad choice. But it refused to take no for an answer. At one point it wrote TypeScript into JSX files after renaming them from .tsx.

I remember thinking: you are incinerating my tokens. Please, help me save the tokens.

The name was born.

Then life happened — family, kids, work, sleep, more work — and it’s 2026 and I’m finally building the site: savethetokens.com (And now people token flex.)

In the gap I accumulated prompt snippets and learnings (mostly via scar tissue), built tools, and settled into a workflow. The models and tools have gotten much better. They still aren’t perfect. The space is moving fast.

I’m sharing what I’ve learned because:

• I’d rather not lose it. (Middle-aged dad memory is not what it was.)

• Maybe it helps someone else.

If you take a look, you will find

• some utility tools

• skills

• prompt snippets

• and a link to Michi

  

  道, “the way”

  my approach to building software, embodied in an agent-centric toolkit.

Here’s the core claim behind Just Ideas: a good idea is only good in context. EDM at a club: okay. EDM at a funeral, well, …

That sounds obvious. But sometimes it’s not.

Every idea arrives with luggage (often invisible and unstated): the context that makes it work. Strip that away and even a brilliant idea looks absurd. Keep it, and an idea you’d normally dismiss might be exactly right.

That’s the starting point for Just Ideas — and it changes how you evaluate almost everything.

Sometimes the context is enormous. Some ideas are timeless and work everywhere (like donuts are good). Other times, the context is narrow: this team, this quarter, this specific mess we’re in right now. Neither scope makes the idea better or worse. It just defines where the idea is useful.

This matters when you encounter an idea you don’t like. Before dismissing it, consider that you might not share the other person’s context. You’re not wrong. They’re not wrong. You are just working out a different puzzle.

And more importantly, ideas should change. They should evolve as you learn, as you accumulate experience, as your world changes. An idea that’s three days old and an idea that’s thirty years old can both be challenged. Ideas are not axioms. They are not etched in stone. They’re working hypotheses with varying expiration dates.

Of course, sometimes — even after you’ve considered the context, acknowledged your biases, and granted every possible charitable interpretation — an idea is just crazy.

That happens too. (As my children will remind me, usually in reference to something I just said.) 😜