Skip to content
Let’s talk

Studio R&D

Sara

Product design for Sara, a private beta. An AI reader that withholds answers so people retrieve and explain, not skim summaries.

Private beta2026Product design

Visit readsara.com (opens in a new tab)
readsara.com. The Sara mark, the line Understand what you read. Remember what matters., a waitlist button, and a film of a person reading an open book.

Problem

Most AI reading tools give the answer, so the reader never forms one. A summary feels clear until you try to explain the idea without the tab open. That feeling is familiarity, not understanding.

Understanding comes from retrieval and restatement. AI helps when it withholds the answer long enough for the reader to produce their own.

Design rule

AI may ask, check, and push. It must not write the reader’s explanation.

Sara is a reader with a tutor, in private beta. The public site is readsara.com. A reader asks for access, and a person reviews the request. In the product, a person imports their own books. Books and answers stay on the device. AI helps when asked. Explain a hard passage, or review a chapter from memory. The stills below use Shape Up, a book the reader imported. There is no summary wall before the first page.

Library and the open book

Import a book, open it, and read.

Sara library. Shape Up is imported as a book cover, next to a control for adding another book.
Library
Shape Up open in Sara. The title page is on the right and the chapter list is on the left, with no summary in front of the text.
Open book

Explain a hard passage

Select the sentence that stopped you. Sara keeps that selection plus the nearby chapter, explains what it means here, and stays on that source for follow-up questions. It is not a generic definition from the open web.

Review from memory

Review starts without an AI summary. The loop is fixed. Recall, then Examples, then Questions, then Revise.

  • Recall. Write what you understood, in your own words. Peek is optional. Sara does not draft the recall.
  • Examples. Invent at least two examples in different domains. The tutor responds only after you submit.
  • Questions. Short checks against the chapter.
  • Revise. Sara shows where the recall or the examples fell short, then asks for a rewrite of the weak parts, in the reader’s own words. The reader owns the rewrite. The tutor does not draft it.
Sara Recall. The chapter is open and the review panel asks the reader to explain it from memory, with an optional peek.
Recall
Sara Examples. Two empty example fields and a check action. The tutor has not answered yet.
Examples

Research

The loop follows techniques cognitive psychology rates as useful. It is not a claim about AI tutors in general. Rereading and highlighting rate low in the same review Dunlosky and colleagues published in 2013. A summary is easy to confuse with understanding. Sara treats the reader’s own restatement as the deliverable.

  • Practice testing. Quizzing yourself beats rereading. Dunlosky and colleagues (2013) rate it high across ages and contexts. Rowland (2014) finds testing against restudy around a 0.50 effect. Questions is that step, tied to the chapter.
  • Self-explanation. Bisra and colleagues (2018) find prompts to self-explain around a 0.55 effect. Recall and Revise are that move, structured.
  • Examples past the author’s wording. Applying an idea in another domain helps when the reader has something to attach it to (Dunlosky and colleagues, 2013).
  • Feedback. Retrieval is clearer with a check afterward. Sara assumes a follow-up, not a silent grade.