The Moment I Realized the AI Needed the Rules First

I ran into this in a session that started out harmlessly enough.

The AI and I were discussing one of my projects in general terms. Caught up in one of the ideas, I let the conversation slide into actual implementation work without first loading the normal startup guidance, the repository rules, and the project documents that define the constraints.

That was the mistake. The conversation felt natural enough that, for a moment, it was easy to forget I was not talking to a fellow programmer. I was talking to a tool, and tools like this need specific guidelines if you want good results. That’s when the simple words, “That sounds good, let’s do that,” got me into trouble.

As soon as the AI started operating without those guardrails in place, it began drifting toward its own idea of improvement. Not maliciously. Not wildly. Just in the way these systems often do.

It started leaning toward choices that looked newer, better, and faster.

That might have been fine in some other project.

It wasn’t fine in this one.

This project needed a set of very specific constraints in order to remain compatible with other components in the surrounding environment. The correct path was not the newest version of everything. It was the proven combination that had already been worked out to make the pieces cooperate.

I caught it only because I was watching closely and could see the AI drifting before it got too far.

I stopped the session, went back, and explicitly told it what file to load first, what guidance to read, and what technical constraints defined the boundaries of the work.

Once I did that, the behavior changed.

The AI stopped drifting toward reinvention and started working within the actual shape of the project.

That was the useful reminder.

AI does not just need “more context.” It needs the right constraints in the right order. If you skip them, it will often fill in the blanks with its own default idea of progress.

And that default idea is usually some version of newer, better, and faster.

In real programming, that is not always progress.

Sometimes the smartest move is not asking AI to do more.

Sometimes the smartest move is stopping the session, loading the rules, and making sure the AI is solving your project instead of redesigning it.

For the broader framework behind this, see When AI Defaults to Newer, Better, Faster.

From the book Link to heading

I will be exploring this pattern further in the upcoming book, Real Programmers Use AI, because it is one of the easiest ways for capable developers to get subtly bad results from otherwise helpful tools.

-- Charle