Challenge 3: Why AI-Narration Pipelines Are Structurally Limited on
Intent -- Solution Walkthrough

The limitation isn't a matter of the underlying AI model needing to
get smarter or better-trained over time -- it's structural, tied to
what information the pipeline actually has access to in the first
place. A post-hoc AI-narration pipeline (Chapters 6 and 7) generates
its narration from the recorded screen actions alone: clicks,
keystrokes, scrolling, window changes. Those actions are genuinely
all the pipeline has to work with, and they only ever encode WHAT was
done, never WHY it was done that way. The reasoning behind a
decision -- why this approach was chosen over an obvious alternative,
what tradeoff is being accepted, what would break if a step were
skipped -- exists only in the mind of the person who made the
recording, and was never captured as part of the recording itself in
any form the pipeline could read.

Because that "why" information simply isn't present anywhere in the
input the pipeline receives, no amount of improvement to the AI
model's own narration quality can conjure it into existence -- a
better model can describe recorded actions more fluently or naturally,
but it cannot infer a reasoning that was never recorded in the first
place. This is exactly why the chapter frames it as a structural
limitation of the category rather than a current quality gap: closing
it would require the pipeline to receive intent information as an
input somehow (e.g. a script, or annotations provided by the
presenter) rather than simply getting better at describing screen
actions it already has access to.

WHY THIS WORKS AS AN ANSWER
------------------------------
This exercise checks that the reader distinguishes "the AI isn't good
enough yet" from "the AI was never given the information it would
need" -- correctly identifying that the limitation is about what data
the pipeline has access to (recorded actions only), not about the
current capability ceiling of the underlying AI model.