Wortree

The problem

A dense book loses you one page at a time, and looking something up means leaving the book, which is where the thread gets lost. The goal was an explanation that appears on top of the page you are on, remembers what you have read so far, and never tells you what comes next.

How it works

Dashed boxes are external services.

Key decisions

• The explanation opens over the book

  The app is built around an iOS share extension: you share the screenshot and a panel streams the answer on top of Kindle, so you never leave the page. The extension is a separate sandboxed process, so it reads the reader's session and book settings from a keychain group it shares with the app.

• One reply does three jobs

  The same streamed reply carries a hidden line that names the book and another that distils the page into a memory note. Pages file themselves under the right book and the memory grows without extra model calls. A reply that is cut off mid-stream is reported as an error and never filed as a card.

• Memory that reaches back without spoilers

  Each page's note is stored per book, and the four most relevant ones are recalled into the next prompt, so page 90 can pick up an idea planted on page 12. The context comes only from pages the reader has captured, and a page detected as belonging to another book never writes into this one's notes.

• The first word in under two seconds

  At the model's default reasoning level the first word took about 12 s. A local benchmark that runs the same prompt file the app ships showed the lowest level starting in about 1.7 s, so that became the default and depth became a per-book setting. When the provider's model alias moved to a version that rejects that level, every route failed at once; the server now retries one level up on that exact error and remembers it per model.

• Prompt changes start from stored pages

  Before redesigning the reader, 71 stored pages were measured. All of them parsed into clean structure, but 5 opened mid-word and 4 of 307 quotes ended on a hyphen-broken word. Both turned out to be prompt defects and were fixed there, and the reader now lays each card out as separate points with the quote they came from.

Results

38 days

from the first commit to live on the App Store, with sign-in, sync and subscriptions in placeMeasured

1.7 s

to the first word of an explanation at the default depth, against about 12 s at the model's default reasoning levelMeasured

0.8 s

to start a story-so-far recap over 30 page notes, matching the quality of the setting that took 7.8 sMeasured

71 of 71

stored pages parsed into clean point-by-point structure before the structured reader shippedMeasured