How, When and Why We Use AI

1. Software Development. I picked up some basic HTML/CSS years ago and I'm comfortable in Terminal, but I am NOT a software engineer. Instead, I follow a "Spec-Driven Design" (SDD) process, mainly using Claude.ai and Claude Code:

1. I prompt Claude.ai with an idea for a product feature.
2. I chat back and forth with Claude.ai to refine it. Once I'm satisfied the core idea is sound, I have Claude.ai write a specification.
3. I review the spec, clarifying decisions or asking for edits.
4. Once it's ready, I hand it to Claude Code to build. The spec instructs Claude Code to pause at a series of checkpoints, where Claude.ai reviews the work and makes adjustments to keep it matching the spec.
5. I manually test every change to confirm it works as intended.
6. For anything particularly hard, I bring in ChatGPT to independently review the code or spec and challenge Claude.ai — it's genuinely interesting watching two different LLMs argue about a problem.

2. The AI Coach.

Why it doesn't do its own math. Early on, I assumed the Coach could just handle numbers directly — ask it for your projected savings at 65, and it'd calculate it. It took me a while to realise what was actually happening: the model wasn't calculating anything. It was narrating something that sounded like a calculation, confidently, in a way that was often wrong and impossible to catch just by reading the response. That was an uncomfortable thing to discover in a product built partly on financial projections.

The fix took a lot of testing, but the principle turned out to be simple: never let the AI do arithmetic it can get subtly wrong. So we built a full library of deterministic functions — real code, not a language model — for every calculation the Coach might need: your Wealthspan projection, your savings rate, your financial independence age, all of it. (The Wealthspan section of our Methodology walks through the actual formulas — that's exactly what the Coach is calling.) The Coach's job is to figure out which function to call and explain the result in plain language. The math itself never touches the AI.

Why patterns are detected by rules, not by asking the AI. This one came out of a conversation with Claude, not a decision I made on my own. I wanted the Coach to notice things — a declining sleep trend, stress rising alongside it — and my first instinct was to have the AI look at the data and figure out what stood out. Claude pushed back: fixed, deterministic rules would catch the same patterns far more reliably than asking a language model to eyeball trends and guess. It was right. So Pattern Detection runs as its own system, entirely separate from the Coach — the same logic, every time, for every user — and the AI only ever explains findings it's handed, never invents its own.

I think that's a genuinely good example of what this whole process looks like at its best: not me directing an AI, and not an AI just doing what I asked, but an actual back-and-forth that landed somewhere better than either of us would have reached alone.

3. Learn / Blog Articles. These are foundational knowledge pieces — the benefits of cardio and strength training, the impact of fees on portfolio performance — and they're >95% written by Claude.ai, with every one reviewed and edited by me. This content is uncontroversial and well-established; the AI isn't forming an opinion or attempting new research, it's summarising existing knowledge, which is exactly what these models are best at. The value isn't novelty — we're not claiming to break new ground by explaining why VO2 max matters — it's having one reliable, curated source for the topics that matter most to longevity.

4. Founders Journal Articles. Unlike the Learn/Blog articles, these carry a personal opinion or story, and that's best written by an actual human — me. Claude.ai does at most a light editing pass. This is where you'll find the thinking behind the platform — pieces like "Why I Built 100 Great Years." These articles are tangential at best to longevity; you don't need to read them to live a longer, healthier, more financially secure life. But hopefully some are interesting, maybe even entertaining (there's an "AI Bloopers" article coming), and they put a human face on the platform.

5. Website Home Page. Roughly 75% me, 25% AI — mainly Claude.ai, with ChatGPT brought in to challenge our decisions. I've obsessed over this website: edited it more times than I can count, once prompting Claude.ai for changes from my phone on the bullet train in Japan. It's the public display of my vision, and I'm a mostly left-brain person — logical, reasoning, math — who knows creativity is my weak spot. Building something that looks good and reads well has been harder than the entire Spec-Driven Design process. I doubt I'll ever call it finished, but I'll keep pushing it closer.

6. Images. Over 99% AI-generated — prompts written with ChatGPT, images generated with Gemini (the photo of me holding my newborn son is real). These images make the Blog and Learn articles more visually engaging and help convey the feeling behind 100 Great Years — living deliberately, feeling secure, being free — but they don't add much beyond that. AI was simply the fastest way to produce that many images.