The Three Ms
of AI™

Here's the core habit: before doing any task the old way, ask yourself — "How could AI do this?" And if the answer is "it can't do all of it," the follow-up is: "How could AI assist with the first 30%?"

It's never binary. The real question is always "to what extent can AI be leveraged here?" Maybe it handles 80%. Maybe 10%. But you don't know until you ask.

Real example: I needed to update tracking links across 300+ YouTube video descriptions. The old me would have opened each video in YouTube Studio one by one. Instead, I described the problem to Claude Code and walked to the kitchen. By the time I came back, it had researched the YouTube Data API, figured out the quota limits, written a script, and laid out a plan. I approved it, it ran, and now I have a reusable system.

The Default Shift is like learning to type instead of writing by hand. Once it clicks, you physically cannot go back. Every manual task starts to itch.

One important thing: AI is better than you think, and improving faster than you think. If AI can't do something today, try again next month. Seriously.

Your job description probably has about five bullet points. Each breaks into dozens of tiny tasks. You don't automate your whole job. You automate one tiny piece. Then another. Then you chain them together.

Real example: Think about "automating a YouTube video." Sounds impossible, right? But break it down: ideation, scripting, title generation, thumbnail generation, description writing, comment replies, timestamps, analytics. Each piece is its own automation. Build one, get it working, move to the next. Chain them together over time and suddenly you've automated a massive chunk of a process that seemed untouchable.

One small task per day. That's it. Do that for six months and you've automated hundreds of tasks. The compounding is real.

This one separates the people who actually get good at AI from the people who just use it like a fancy search engine.

Never accept AI output without asking why. Ask for three alternatives. Ask which one it thinks is best and why. Push back. Dig in.

This is the antidote to what I call "dark code" — automations or code that you don't understand. If you build something and you can't explain how it works, you've built a liability, not an asset. When it breaks (and it will), you'll have no idea where to start.

If you're not curious, you're just pushing buttons. Every question you ask makes you smarter for next time. Treat AI as a mentor, not a vending machine. The vending machine approach gives you output. The mentor approach gives you understanding.

Break your automation into the smallest possible steps. One input, one output per block. The output of block 1 becomes the input of block 2.

Start with zero-AI steps first. Get the deterministic pieces working — the data fetching, the formatting, the routing. Then layer in AI where it's actually needed.

This does two things. First, it makes the project less overwhelming. Second, it lets you validate as you go. If block 3 is producing garbage, you know exactly where to look.

Think of everything as Lego blocks that can plug into other systems. Modularity is freedom.

Like workers on an assembly line — each person has one task they're great at.

Don't build a generalist. One model call for writing copy. Another for reasoning through a decision. Another for classifying input into categories. Keep them separate.

Benefits: easier to debug (you know which step broke), easier to swap models (upgrade one without touching others), easier to adjust prompts (tweaking the classification prompt won't accidentally break your copywriting step).

Validate each step's output before chaining it to the next. Do NOT build the whole pipeline and then test end-to-end. That's a recipe for "it doesn't work and I have no idea why."

Build step 1. Run it. Confirm the output is correct. Build step 2. Run it with step 1's actual output. Confirm. Chain them. Add step 3. This is how you build a proof of concept that actually works.

Every time you skip validation and chain three steps together untested, you're creating a debugging nightmare for future you.

There's no such thing as a finished product — especially with AI.

Deterministic scripts CAN be done. A script that reformats a CSV does one job, does it well, and you can call it finished. But AI steps are always evolving. New models drop. New capabilities emerge. Things get faster and cheaper. The prompt that was optimal six months ago might be verbose and expensive today.

Ship the POC. Get feedback from real usage. Expand. Iterate. Then iterate again. Perfectionism is the enemy of deployment.