Why does an AI summary of a long transcript cover the beginning well and the end badly?

Attention thins across a long input, and the end is where it thins most, so the symptom is a smooth gradient from a detailed opening to a compressed tail rather than scattered errors. Splitting the transcript into four parts, outlining each separately and then merging restores the ending — which matters because meetings often reach their real decision in the last ten minutes.

Why does asking for every theme in a set of interviews return dozens of them?

Material gets worked through occurrence by occurrence rather than by distinct thing, so a single theme returns under four different names. Ask for each distinct theme once with the number of mentions beside it: forty collapses to roughly nine, and the mention count is a finding in itself, because the theme raised nine times is usually the one that matters.

What is the difference between a topic and a claim?

A claim can be false and a topic cannot. The economics of remote work is a container that commits to nothing; remote work costs more in coordination than it saves in rent is a position somebody can argue against, and naming the alternative is what gives every section afterwards a job to do.

From source to structure

Nothing on this page calls an AI. You run everything in your own account.