Before you touch Claude or Gemini, ask a harder question: is this actually going to move the business forward, or am I just chasing a shiny idea?
Most operators have one to three repetitive tasks that can be handled safely and reliably by AI. Those are the ones worth automating. If you're not bleeding time or money on manual work, automating for the sake of it is a waste.
The best candidate for AI is something you already do today. Not an idea. Not a hunch. A live, running process. If you're a one-person team doing everything, the rules bend — AI is how you scale yourself. But if you're running a real business and the process doesn't exist yet, build it manually first, then automate.
Here's the whole video in one line: log in to your AI, describe what you want in ultra-specific detail, and tell it to treat you like a sixth grader in that field.
That's it. Ultra-specific end goal. Ultra-specific constraints. Ultra-specific outcome. Then ask it to walk you through the build one step at a time as if you know nothing.
This is how I built my first live-feed dashboard without ever having built a dashboard before. It's how I put together my content auto-publisher and scheduler. I described the dream outcome in painful detail, told the model to teach me like a beginner, and stayed in the driver's seat the whole way.
Vague inputs are how you end up in the horror stories — servers wiped, databases cleared, agents going rogue. Those aren't AI problems. Those are limitation problems. If you don't tell the model where the fence is, it will run past it.
You don't need any special credentials to use AI well. If you can sit down, focus for a couple of hours, and finish something, you already have 90% of what's required. The rest is mental.
**Articulation.** The ability to describe what you want in ultra-specific detail. A vague prompt gives you twenty variations of the wrong answer.
**Critical thinking.** AI is built on what humans posted to the internet. It has no real-world experience. You do. Don't blindly accept the first answer — push back when it feels off.
**Cause and effect.** The model doesn't know your business. It can't see how a system might break something downstream. That's your job.
**Patience.** Building takes hours. Your patience will get tested. The upside: you can tell the model it's underperforming and it will often course-correct. Some LLMs literally detect frustration and shift approach.
**Nuanced comprehension.** Understand how each piece connects. The chat-logs-indexed-to-Google story happened because people used the tool without understanding what it was doing. Treat AI as a tool, not an autonomous employee.
The personal side is how you talk to the model. The business side is how you decide what to build in the first place.
**Bottleneck analysis.** Where is the constraint? Automating a non-bottleneck is just faster noise.
**Data-driven decision making.** Identify what actually 2x's the result. Then act on that — not the fun idea.
**Speed of execution.** Move. Delegation is baked into agentic AI itself — a master agent hands work to sub-agents. If you can't delegate to humans, you can at least delegate to the system.
**Risk tolerance.** It won't work the first time. Or the fifth. I've spent a lot of hours learning that things take longer than expected. Lengthen your time horizon and success becomes inevitable.
Once you know what you're automating, decide what type of build it is. This one decision changes everything downstream.
**Deterministic** is a fixed, hard-coded setup that produces the same outcome every time. Inputs vary — a form submission, a lead, a file — but the output shape stays constant. It gets routed to a receptionist, a setter, a sales rep. Predictable. Boring. Reliable.
**Agentic** is looser. Still controlled when scoped right, but the model uses reasoning to dynamically pick which steps or tools to hit to reach the outcome. This is where you get multi-tool workflows, sub-agents, and dynamic decision-making.
Repetitive, time-draining, cash-draining tasks — those are usually deterministic wins. Bigger builds where you're bringing software in-house to kill a subscription — those tend to lean agentic.
Once you've picked, the rest of the build is knowledge base, prompts, rules, triggers, and API access. But none of that matters if you haven't nailed the first four steps.