It will take months to build! So...?

This post is something that developed from my win post last week and something

has mentioned a couple times in his last teas were he is talking about a project he is doing for a customer.

What started as a one month website rebuild is turning into a 4-6 month project. And its the best thing that could have happened.

Original job was simple. Build a new site for work using ICM to pull it together. But when I started gathering product info to display to customers in a structured way, I realized I had 2/3rds of our product data sitting in one place instead of 23 different locations for the first time in the company's history. So instead of grabbing what I needed and moving on, I stopped.

Why not build a real source of truth while I'm in here?

Two files per product. One with the hard data, specs, features, the stuff that doesn't change much. The second with images and videos linked, copywriting notes, all the context that makes things actually useful. Once those existed, everything else opened up. Manuals, datasheets, social posts, internal training docs. All of it buildable from the same foundation of knowledge/information/truth. The ICM grew from there. More data in, more use cases out.

A month long project is now 4-6 months. And I'm completely fine with that.

The best part, when we make a new product? Fill out the 2 templates, drop it in to the SoT directory, tell the AI there is new files, it updates the structure and we're ready to go.

Here's the thing I keep coming back to.

There's a pull with AI to rush. You can build a folder-based agent in 30 minutes, so the feeling is: why aren't you done? That feeling is a trap.

Fast to build does not mean it works. Or that its good. Or that it does what a customer actually needs when they're sitting in front of it for real.

Jake's 60/30/10 rule for designing a folder based agent is 60% infrastructure, 30% orchestration, 10% AI. That ratio doesn't stop at the agent design. It applies to the whole project.

Spend 60% of your time gathering the data. 30% organizing it. (or even 30% gathering, 60% organizing. Either way, the bulk of your time is spent here) Then 10% building the agent. The conversation with the AI about folder structure should be the last thing you do, not the first. Have as much squared away before you even open the IDE as you possibly can.

This is something important to remember when you build for someone else. What quirks you are willing to tolerate with a build for you, won't fly with the customer. So the more work you do upfront, the better the output will be in the end.

Garbage In, Garbage Out. Its been a mantra in the software world for at least 50 years now and it even more true today in the Age of AI as it was back when code was still on written on punch cards.

We've spent 20 years building domain knowledge. The products, the customers, the workflows, all the context that makes any of this useful to someone other than me. A machine didn't give me that. Experience did, and most of it learned the hard way.

4-6 months to turn that into something that actually is actually useful and runs is not slow. Its nothing, when we took 20 years to get to this point.

Build it right from the beginning. What are you rushing toward anyway?

3 comments