The naive version summarises once and reformats the summary nine times, which produces nine assets that all lead with the same sentence in different aspect ratios. Reach does not move, because nothing new was said.
The version that works decomposes first: read the source and inventory every distinct claim, with its evidence and how surprising it is. A decent long article yields eight to twenty. Then match different claims to different formats, because formats have appetites.
A carousel wants a counter-intuitive claim with three supports. A quote card wants a sentence that survives without context. An answer-engine snippet wants the definitional claim stated plainly. They are different jobs and they want different raw material.
Why do repurposed assets all feel the same?
Because they were generated from one summary, so they inherit the same compression and lead with the same claim. Inventorying the claims first is what makes the outputs genuinely different.
How many usable claims does a long article contain?
Typically eight to twenty in a 1,500-word piece, which is more than enough to give nine assets nine different openings.