Blog
Notes from building an on-device reading tool — what the models actually do, what breaks, and which parts turn out to be harder than the part everyone expects.
A recap covers one flick, not everything above it. Measured coverage from 1,200 to 9,000 px, where it stops being complete, why the points stay at seven anyway, and the measurement run that had to be thrown away.
Feynman, a recap that stops at three points instead of ten, a retry that actually rephrases, sentence separators chosen by language, and a card the host page can no longer restyle.
What triggers a recap, what the popup controls, how the panel reaches the cursor, and why the two layouts come from a single generation rather than two.
Measurements from one 72-generation Skim Recap run, including prompt configuration, output lengths, timing records, language probes, artifact boundaries, and open questions.