Blog

A few days ago I read a forum post by someone furious with an AI assistant. It had "forgotten" something they had told it. Part of me sympathized. AI can be frustrating that way. Another part of me, though, took the side of the AI. Let me explain why.

OTS Templates only work if the AI assistant has the right knowledge in front of it at the moment it does each piece of work. This post is about two ways that goes wrong. One was a plumbing problem which we fixed. The other was a writing problem; that one was mine. I'm still learning how best to leverage the power of AI.

The library gets its first public showing at two upcoming AUG chapter meetings:

• Pacific AUG — October 1, 2026
• Europe AUG — October 7, 2026

A stubborn error in one of my templates led to understanding about Domain Aggregate expressions inside transactions in VBA. I'm sharing that with you in case it is something you might find useful some day.

In the previous blog, I talked about the four builds that followed different paths to reach the same destination. It was largely a success story.

I would be untruthful, though, if I didn't also share the negative side of the ledger: the deviations and false starts exposed in the build report and runbook. They happen. We learn from them and move forward. But there's also a good story to tell about that.

A few days ago I ran the same audit-logging template four times against Northwind-derived test databases. As I expected, no two runs built the code the same way. That is not a bug in the template. It's AI, so I expect that.

The thing is that I had put a lot of work into crafting specifications and validation checks to make it work correctly despite those variations. I was very pleased, therefore, because even though the build paths were diverse, one thing didn't change. All four versions logged records in the audit table the same way.

Today's blog is a response to Mike Wolfe's "Outcome-Based Instructions."

His argument in one line: with today's more capable AI models, tell them what you want to end up with, not how to get there.

Several people have asked me, "What does a template-driven build actually give me that telling an AI assistant to do something, like add audit logging to this database, does not?"

And frankly, the answer to that question, once I really thought about, was not what I had expected it to be.

The real advantage, I believe, is asking the right questions.

I thought I had been careful enough. The AI Agent proved me wrong. We almost broke up over it. We've reconciled, but the trust problem lingers.

By the way, I'm not telling this story to get your sympathy. I'm sharing it so you can possibly avoid the mistakes I made.

Let's make this a dialog, not just a billboard.

After getting a request from a reader to allow feedback, I had Claude Code add the ability to do that. Let me know what you think, good or bad. Especially if it's good, though.