How to Build a Simple AI Workflow From Scratch

When I first started hearing people talk about their "AI workflows" I assumed they were talking about something technical that I wasn't qualified to do. Some complicated pipeline involving APIs and code and things I didn't understand. So I ignored it for a while and just kept doing things manually like a person in 2019.

Then one day I sat down and actually tried to build one, just to see. And honestly it was fine? It took me an afternoon. And now there are several recurring tasks in my life that I just... don't do manually anymore. I want to explain how I think about this, because I think the concept gets way over-complicated in the way people talk about it.

What an AI Workflow Actually Is

A workflow is just a sequence of steps that you repeat regularly. That's it. An AI workflow is when one or more of those steps involves an AI tool doing some of the work.

You probably already have workflows you don't think of as workflows. Every time you create a new piece of content, you probably do roughly the same steps in the same order. Every time you answer a certain type of email, you follow a pattern. Every time you research something, you hit roughly the same sources.

Building an AI workflow means looking at those patterns and figuring out where AI can do a step faster, better, or both. And then making that the default rather than a thing you have to remember to try.

Step One: Pick One Repeating Task

Don't start big. I made that mistake initially, trying to automate my entire content process at once, and it got overwhelming and I abandoned it.

Start with one specific task you do regularly that has consistent inputs and outputs. For me, the first thing I automated was writing product descriptions for my Etsy shop. I was doing roughly the same thing every time: taking notes about a new product, writing a description that hit certain keywords, formatting it a particular way.

That's a perfect candidate. Same type of input every time (product notes). Same type of output every time (a formatted description). Clear quality criteria I could evaluate against.

Step Two: Write Down What You Currently Do

Before you involve AI at all, write down every step in your current process. Like actually write it down, even if it feels obvious.

For my product descriptions, the steps were: look at my product notes, identify the main features, think about who would buy this, write a first draft, edit for keywords, format with line breaks and a call to action, copy into Etsy.

This matters because you need to know what you're automating before you can automate it. And often when you write it out you realize there are more steps than you thought, and some of them are actually pretty mechanical and easy to hand off.

Step Three: Figure Out What AI Can Take Over

Go through your step list and for each one ask: is this a thinking/judgment step or a mechanical/production step? Judgment steps (what angle do I take on this, is this quality good enough, does this fit my brand) stay with you. Production steps (turn these notes into a draft, format this text, generate options to choose from) are candidates for AI.

In my product description workflow, almost everything after "think about who would buy this" became an AI step. I'd do the thinking, give the AI the output of that thinking, and let it handle the drafting, keyword work, and formatting.

The quality judgment, "does this actually sound like me, is this good", that stayed mine. You don't outsource that part.

Step Four: Build the Prompt

This is the actual work. You're building a prompt that captures all the context and requirements that used to live in your head.

I started by writing out all the things I was implicitly doing when I wrote a product description. Tone: warm but not gushing. Keywords: I have a list of about fifteen I rotate through. Structure: start with the sensory experience, then practical details, then a brief call to action. Length: 150-200 words. Things I never include: discount language, excessive exclamation points, hollow words like "beautiful" without specifics.

That all went into a prompt template. And then I tested it against some products I'd already written descriptions for, comparing what I'd written to what the AI produced. Where they diverged, I figured out why and adjusted the prompt.

I went through probably eight or ten iterations before I had a prompt that was consistently producing stuff I'd actually use. This part takes time. It's not magic from iteration one.

Step Five: Make It Easy to Actually Use

This is the step most people skip and then wonder why they're not using their workflow. If accessing your workflow requires more than about thirty seconds of setup, you're going to keep doing things the manual way when you're busy or tired.

I keep my prompt templates in a single doc. When I have a new product, I open the doc, copy the template, fill in the product-specific blanks (takes about two minutes), paste it into Claude, and I'm done. The whole thing is maybe five minutes instead of the thirty it used to take.

Other options: some people save prompts in their AI tool if it has that feature. Some people use a simple notes app. Some people build this into something like n8n if they want it more automated. All fine. The point is that it should be frictionless to access and use.

How to Expand Once One Workflow Is Running

After my product description workflow was running smoothly. And I mean actually running, not just theoretically working. I looked for the next thing. Not before.

I added a workflow for responding to customer questions. Then one for generating content ideas from my analytics. Then one for outlining blog posts. Each time, same process: write down the steps, identify what AI can take, build the prompt, iterate, make it easy.

Now I have five or six of these running and they genuinely save me several hours a week. Not because I did something complicated, but because I was systematic about it.

You don't need to be technical to do this. You need to be willing to think clearly about what you actually do, and patient enough to iterate on a prompt until it's good. That's it. I promise the first one is the hardest.