Skim Recap

Detects fast scrolling and summarizes the skipped text on-device with Gemma (LiteRT-LM).

Skim Recap summarises the part of an article you scrolled past — on your own machine. No page text ever leaves your browser. Built by Yui Morii — https://skim-recap.vercel.app When you flick through a long article, documentation page, or essay, Skim Recap notices the passage you skipped and shows a short recap beside your cursor. You stay on the page, keep your place, and get the context back without scrolling up. Each recap covers that one flick, not everything above it. Flick to 4,000 px and you get those 4,000 px; flick again to 8,000 and the next recap is of the new stretch, not of the whole 8,000. The pixel count on the card is that distance. Summaries are generated locally by Gemma 4, running on your GPU through LiteRT-LM. There is no API key, no account, and no hosted summarisation server: article text is never sent anywhere for summarisation. BEFORE YOU INSTALL Running a real language model in the browser is not free. On first use, Skim Recap downloads a 2.97 GB model, which is then cached locally for later sessions. You will need WebGPU support, about 22 GB of free disk space, and either more than 4 GB of video memory or 16 GB of system memory. If that does not describe your machine, this extension may not be a good fit yet — better to know before installing. WHAT IT DOES • Detects a fast scroll and recaps only the passage you skipped • Explains a term the page never defined: select any word or phrase — on the page or inside the recap itself — and press Explain. A recap is bounded by the passage on purpose, so it can name a term three times without ever saying what it is; this lifts that bound for the phrase you picked. It never runs unasked • Ignores ads, navigation, comment threads, and "related articles" sections by scoring page blocks using signals such as link density, element class names, and content length • Excludes content you already read, so the recap focuses on what was skipped • Samples content across the skipped range, rather than relying only on its opening paragraph • Shows the recap in a floating panel beside your cursor • Scales the number of points to how much you skipped — 3, 5 or 7 — and drops a point that repeats one already made • Includes two layouts, Focus and Smart, which are two readings of a single generation rather than two summaries: Smart leads each point with a short label so the points can be scanned, Focus drops the labels and runs them together as prose. Switching does not re-run the model • Retries by asking for a different way in, so a second press gives you different words rather than the same ones again • Generates the recap directly in a language you choose, independent of the page language — for example, a Japanese article can be recapped directly in English • Translates a finished recap on demand using Chrome's on-device translation. The button names the language it will give you, and pressing it again brings you back • Lets you copy the recap in whichever layout is currently active • Includes an adjustable flick threshold • Supports an optional keyboard shortcut to switch Skim Recap on and off, configurable at chrome://extensions/shortcuts • Includes a model preload button with download progress Because recap text is generated by a language model, wording and length may vary between generations. WHAT IT DOES NOT DO YET • Search-results pages do not produce a recap. Result titles are links and the snippets beneath them are ordinary prose, so what came out was a recap of the search rather than of anything read • Pages that hold their prose in generic containers rather than paragraphs, as some editors emit, may not be extracted at all • A very long single scroll — several screens without pausing — is sampled rather than covered completely. One or two ordinary flicks are covered in full PRIVACY Skim Recap reads page text only to determine what you skipped and to generate the recap. That processing happens locally in your browser. The model is downloaded from Hugging Face when needed. Article text used for summarisation is not transmitted to a hosted AI API, logged remotely, or stored off your machine by Skim Recap.