The method

What the tool does, and what it cannot know

Two modes sit behind one desk. Passage mode picks sentences out of your text without a model. AI mode writes points and must cite the words behind them.

AI mode, step by step

Your text is split into numbered paragraphs in the same way as passage mode. Those paragraphs, your length choice, and any emphasised paragraph are sent once to an external language-model service, which is asked to return a JSON list of points, each with a short summary, a quote, a paragraph number, and a kind.

Two independent checks run before anything reaches the page. The service's own parser rejects a point unless its quote appears verbatim inside the paragraph it cites. The page then re-locates each quote against the copy of the text you still hold in the browser, and discards any point whose quote appears zero times or more than once. Only points that pass both checks are shown, with a note telling you how many were removed and why.

Every displayed point is labelled fact, opinion, or inference. The label describes the kind of claim in the text, not whether the claim is true. Nothing on this page checks the truth of a sentence.

What the model cannot fix

A verified quote proves the sentence exists in your text. It does not prove the point is important, complete, or correctly labelled, and it does not protect you from a passage that is true out of context. Read around each highlighted span before you quote, paraphrase, or act on it.

Failure and cancellation

If the model does not answer in time, returns unreadable output, or returns only points that fail verification, the page says so plainly and shows no summary. It never substitutes a fixed or stock summary. Cancel stops the request; the service may retry a failed call once, and a cancelled request is not retried.

Selection, step by step

The tool divides pasted plain text into paragraphs and sentences in your browser. It counts recurring meaningful English words, gives a small preference to sentences at the start of a paragraph, and distributes selections across paragraphs where it can. A paragraph you choose as a focus receives extra weight. Brief, balanced, and detailed show up to 3, 6, or 10 sentences.

Selected sentences return in their original reading order. A result displays its paragraph and start/end character offsets; the end position is exclusive. The source view highlights that exact span. Offsets are JavaScript string positions, so emoji and some other characters can count as two units.

Passage mode is not a semantic summary

Word frequency and position do not reveal whether a claim is central, credible, or representative. A short selection can omit a qualification elsewhere in the document. Passage mode does not rewrite the author and does not classify facts, opinions, or inferences; use AI mode when you want labelled points, and accept that it is a model's reading rather than a verified one.

Phrase cues are only wording cues

The tool marks phrases such as “I think”, “in my opinion”, “should”, “will”, “may”, and “might” when they occur in a selected sentence. Those expressions can suggest opinion, recommendation, or uncertainty. They do not prove the category of a statement and do not tell you whether it is true.

Input and failure

Only pasted plain text is supported, up to 120,000 JavaScript characters overall and 4,000 per sentence. PDF, DOCX, URL import, and file upload are not available. An over-limit input is rejected in full, with its actual character count shown. Empty input gets a prompt to paste text. If the tab is refreshed, text and selections disappear.

Open the reading desk or read the source checking guide.