The map and the path

Being AI-proficient often means knowing what AI is good at, what it is bad at, and using it accordingly.

At its core, an LLM is a very smart autocomplete machine. It predicts the next word based on the context window it keeps in "memory."

Intuitively, this makes you think it should be good at connecting the dots.

You give it a problem and a desired destination, and it tries to find a path.

On the first iteration, it will usually find the obvious path. The generic path. The averaged-out path that fits some hypothetical situation.

But your situation is not generic. Your situation has context.

So you start adding more points to the map: milestones the path should go through.

The path becomes more real.

But then you are still not satisfied with it. And of course you are not.

Because so far, you have only defined what you want in.

Now you need to define what you want out: the constraints, the borders, the things the solution must not do.

As you add these dots and borders, the path the LLM draws gets cleaner and better.

Eventually, you might even start liking what you see.

But then, after adding the 11th milestone, you suddenly notice that the new path skips milestone #2. Or ignores one of the constraints you clearly outlined.

You say: "Fix it."

Sometimes it works.

Sometimes it does not.

Sometimes the whole thing becomes even more messy.

So what happened?

In my experience, it is usually one of two things:

1. The model cannot keep all the important points in working memory, or it does not understand what is truly important versus what was just filler.
2. There is no valid path through the map you created. The task cannot satisfy all your dots and constraints at the same time, even if you really, really want it to.

And it is surprisingly easy to distinguish between the two:

Try doing it yourself.

If you cannot draw the path manually, the problem is probably not the model. The inputs need to be redesigned.

But if you can draw the path yourself, and the model keeps losing parts of it, then the issue is usually context management.

For the first category, you drop the impossible combination and redesign the task.

For the second category, the solution is usually to re-emphasize the important milestones and constraints, or to zoom in and out of the map so the model is working with the right level of context.

This is where the real skill begins.

Not prompting.

Not "talking to AI."

But learning how to structure context, constraints, milestones, memory, and working scope.

Thanks to tools like Claude Code, we now have skills, agents, and workflows for this.

And I think these are becoming the crucial skills to master.

Without them, LLM work does not scale beyond simple isolated tasks.