Summarise a chapter, then find out what the summary threw away

WeSolve+ is best for a student who summarises a chapter and then wants to be tested on it, because a summary you only read is the weakest thing you can do with the hour. What the tool cannot do is open your file or write in its own words, which is the part the app handles.

A summary is a map, and you cannot revise from a map

Shortening a chapter tells you where things are. It does not put them in your head, and the evidence on this is not close: in the review by Dunlosky and colleagues, summarisation and highlighting were both rated low utility, while practice testing and distributed practice were rated high. Summarising is genuinely useful for one job, which is deciding what to spend the next hour on. Treat the output above as a table of contents with opinions, then go and be tested on the parts it kept. The reason it feels productive is worth naming: rereading a condensed version produces fluency, and fluency is mistaken for knowledge, an error close enough to the illusion of explanatory depth that the same correction works, which is trying to produce the material without looking.

Extractive and abstractive, and why this one is extractive

There are two families. An extractive summariser selects sentences that already exist and prints them unchanged. An abstractive one writes new sentences that say what the source said, which is what a language model does. The trade is exact: extraction can never invent a claim, but it reads like clipped highlights and it inherits any bad sentence it picks. Abstraction reads smoothly and can compress two paragraphs into one line, but it can also state something the source never said, the failure known as hallucination. A browser tool with no model in it has to be extractive, and that is the honest option here rather than a limitation being spun.

How the ranking works, in full

The method is old and simple enough to check by hand. Words that repeat across a passage are treated as the subject of the passage, an idea due to Hans Peter Luhn in 1958 and still the backbone of automatic summarisation. Very common words carry no subject matter, so a stop word list removes them before counting, and anything under four letters is dropped as well. Each sentence scores the sum of its distinct content word frequencies, divided by the square root of its length so that a long sentence cannot win on bulk alone. The three words that earned each sentence its place are shown under it, which is the part most tools hide. Weighting by rarity as well as frequency, the tf-idf refinement, needs a corpus to compare against and there is no corpus in a browser tab.

How short is short enough

One sentence in five is the default here because it is roughly the ratio at which a passage still reads as an argument rather than a list. Published abstracts sit far tighter, often 250 words for a paper of 6,000, near four per cent, but an abstract is written by someone who understood the paper and can restate it. Selection cannot compress that hard without losing the connective tissue. A practical rule: if your summary is under a tenth of the source you are almost certainly reading a list of topics, and if it is over a third you have not made a decision about anything. Reading time is the other check, at roughly 200 to 260 words a minute for adults reading non fiction on screen, so a summary you cannot get through in two minutes is not doing its job.

Auditing a summary you did not write

Any summary, from a tool or a model, needs three checks before you revise from it. First, every number and name in the summary should appear in the source, verbatim, and this is where abstractive output fails most often. Second, anything the source hedged should still be hedged; a sentence that turns may into does has changed the claim. Third, look at what is missing rather than what is present, because a selection method skews toward whatever the author repeated, and an author repeats what they find interesting, not necessarily what is examinable. Researchers score machine summaries against human ones with ROUGE, which measures overlapping word sequences; that is a useful reminder that overlap is a proxy and no metric asks whether the summary is true.

Turning the summary into something that tests you

The whole point of shortening a chapter is to know which parts to work on. A summary is also the natural script for something you listen to later, which is what the podcast generator page is about. Switch the output type above to cards or questions and the same pasted text is converted rather than condensed, which takes the summary from a reading task to a producing one. Writing the question yourself is better still: the generation effect is the finding that material you produce is remembered better than material you read, and the older levels of processing work points the same way. Two structured methods already build this in: SQ3R asks you to turn each heading into a question before you read it, and Cornell notes reserve a margin for the questions your notes answer. Both predate every AI tool on the market and both still work.

What the app adds, and what it costs

Upload a PDF, a photograph of a page or a camera shot and WeSolve+ reads the file itself, which no browser rule can do, then writes an explained quiz, a card deck and an audio recap of roughly one minute from that single file. Wrong answers collect in their own list so the next session starts with them. It runs at wesolveapp.com in any browser and as an App Store app on iPhone and iPad. Free to start, with two quizzes and two uploads a day, two recaps a week and two duels a day. There is no separate Android app. Numbers on pricing.

Last updated: 2026-08-15