TD;DW

Too Distracted, Didn't Watch. Spoiler-free "what did I miss?" for the video you're already watching.

Activate it mid-video and you get a recap of everything that happened up to your current timestamp, and nothing past it. Works on YouTube, Netflix, Prime Video, Disney+, and your own Jellyfin server. There's also a follow-up question box, so when the recap leaves you going "wait, who's that guy again?" you can just ask, without leaving the page and losing your spot.

I built this because I kept pausing shows to go "hang on, what did I miss," then falling into a twenty-minute Wikipedia hole and forgetting what I was watching. Now I hit a shortcut and get on with it. Full disclosure: it's a joint project with Claude (Anthropic's AI), and I'm happy to say that out loud.

Left to its own: a spoiler-free recap up to where you paused. Ask a follow-up and it answers right in the panel.

Before you start (read this first)

• You need an LLM endpoint. Pick a provider in the extension options:
  • OpenRouter: one key, every model.
  • Google Gemini (AI Studio): uses your AI Studio key via Google's OpenAI-compatible endpoint. Handy if you've got AI Studio credits sitting there.
  • Custom / Local: any OpenAI-compatible endpoint (Ollama, LM Studio, llama.cpp).
• Model choice matters most in knowledge mode. For titles TD;DW recaps from the model's own memory (Netflix/Disney, or a Jellyfin item with no subtitles), small models hand you vague, plot-shaped filler, so reach for a stronger model there. Transcript mode (YouTube captions, Jellyfin subtitles) feeds the model the real dialogue, so it stays accurate even on cheap, fast models and even for brand-new or obscure content.
• Every recap tells you where it came from. A badge at the top reads either green, "Built from this title's subtitles" (grounded, trustworthy) or amber, "From the model's memory, may be imperfect" (recalled from training, so treat it as a refresher). It's an honest signal, not a made-up confidence score.
• The keyboard shortcut can silently collide with an existing binding, so check or change it at chrome://extensions/shortcuts.
• Disney+ (ad tier): playback position is read from the player's own scrub bar, because the raw video stream has stitched-in ads. If you get "couldn't read the playback position," wave your mouse over the player so the controls appear, then hit "Refresh at current time".

Install

1. chrome://extensions → enable Developer mode → Load unpacked → pick this folder.
2. Open the extension's Details → Extension options, choose your provider, paste your API key (or local endpoint URL), Test connection, Save.

Use

• On a supported watch page: click the toolbar icon or press Alt+Shift+U.
• The video pauses (you can turn that off) and an overlay shows bullet points covering everything up to the current time. In transcript mode, click a bullet's time chip to jump the video to that moment.
• Got a question about the recap? Type it into the follow-up box at the bottom of the panel and ask away. It remembers the back-and-forth while the panel is open and forgets it the moment you close it, so nothing is saved. Answers stay spoiler-free by default, same as the recap; if you'd rather they fill in the whole picture (even past where you are), turn off "Keep follow-up answers spoiler-free" in options.
• Keep watching, then hit Refresh at current time to extend the recap.

Jellyfin

Point TD;DW at your Jellyfin web UI in options (Jellyfin server URL, e.g. http://jellyfin.local:8096) and reload the Jellyfin tab. It runs on the Jellyfin page itself, reads your logged-in session token from the page (no separate API key), and pulls the item's subtitle track up to your timestamp for a transcript-grade recap of anything in your library, movies included. Items with only image-based subs (PGS/VOBSUB) or no subs fall back to knowledge mode automatically.

Permissions & privacy

What it sends, and where. TD;DW has no backend, no analytics, no telemetry. The only things that leave your machine are the recap request (the title, plus in transcript mode the subtitles up to your current timestamp) and any follow-up questions you type along with the recap they're about. All of it goes only to the LLM endpoint you configure, with your own API key. Nothing is logged, and nothing phones home. (The one extra outbound header is an HTTP-Referer/X-Title on OpenRouter requests, which OpenRouter uses for app attribution; it says nothing about you.)

Why the permissions look the way they do:

• storage: saves your settings (provider, model, Jellyfin URL) in chrome.storage.local. Your API key lives here and never leaves except as the Authorization header to your endpoint.
• scripting: registers the Jellyfin adapter on your server's own origin at runtime (that address isn't known at build time, so it can't be a static entry).
• Default site access is deliberately minimal. The extension only auto-runs on YouTube, Netflix, Prime Video, and Disney+ (the content_scripts block in the manifest), plus openrouter.ai for the default provider. That's it on install.
• The broad optional host permission (https://*/*, http://*/*) is not granted on install. It's there so that if you point TD;DW at a custom LLM endpoint or your own Jellyfin server, it can request access to that one origin. You get the normal Chrome prompt for the specific host, nothing wider. Stick to OpenRouter/Gemini on the built-in sites and it's never requested.

Honest caveats:

• In transcript mode your subtitles (and any follow-up questions) are sent to a third-party model; in knowledge mode it's just the title and your questions. Either way, a cloud LLM sees whatever you send it, so pick an endpoint you trust, or run a local one.
• Follow-up questions resend the recap and, in transcript mode, the transcript each time, so a long back-and-forth on a long video costs more tokens than a single recap. A non-issue on cheap models, worth knowing on pricey ones.
• YouTube captions are fetched through YouTube's internal InnerTube API, which is a technique, not an official or supported endpoint, so that path can break when YouTube changes things.

Architecture (for future me)

• Content scripts share globalThis.TDDW (no build step); the manifest loads them in dependency order. The overlay is a closed-shadow-DOM panel that never passes LLM output through innerHTML. (CSS class prefix is cmu-, and the config storage key is catchMeUpConfig; both legacy names, kept on purpose to avoid churn and settings loss.)
• content/adapters/: one adapter per site behind a common interface: matches(location), getVideoState() → {service, title, episodeInfo, currentTimeSeconds, durationSeconds}, and optional getTranscriptUpTo(seconds) → {text, language, autoGenerated, truncated}. Whether getTranscriptUpTo exists is what routes a site into transcript mode. index.js is a first-match registry over matches().
  • Netflix/Prime/Disney share the dom-video-adapter.js factory, so a new streaming site is about 30 lines of config. Site DOM fragility stays in each site's readTitle / readPlayerTime.
  • YouTube fetches captions via the InnerTube ANDROID-client route (web timedtext is proof-of-origin-gated and returns empty bodies). Lines carry [m:ss] prefixes that become seekable chips.
  • Jellyfin (jellyfin.js + jellyfin-subtitles.js) talks to the Jellyfin REST API same-origin: it resolves the playing item, filters to text subtitle streams (IsTextSubtitleStream, skipping image subs), fetches Stream.vtt, and clips cues past the timestamp. It's injected via dynamic chrome.scripting.registerContentScripts keyed off the configured server URL (the static content_scripts block can't know it at build time), and registration is synced on save / install / startup.
• Title fallback chain: site DOM selectors, then Media Session metadata (mirrored from the page's main world by content/main-world/media-session-probe.js, which survives player redesigns), then site-specific extras (Disney uses the tab title; Jellyfin uses item metadata from the API).
• Position: video.currentTime, except on ad-stitched players. The Disney+ ad tier reports duration: Infinity and stream time that includes ads, so those set playerTimeOnly and read the player UI's Timeline slider via TDDW.queryDeepAll(), which pierces open shadow DOM and same-origin iframes.
• background/providers/: recap engines behind pickProvider(). It picks transcript when the adapter supplied captions/subtitles and knowledge otherwise. Each result carries a source field the overlay renders as the provenance badge. Follow-up questions run through buildQuestionMessages in the same folder, reusing the recap (and transcript, when present) as context. Adding an engine is one file plus one registry line.
• background/providers/prompts.js: all the spoiler-safety and identification rules. Iterate here; it is the product.
• background/llm-client.js: a single OpenAI-compatible client ({baseUrl, apiKey, model}), fetched straight from the MV3 service worker, and it strips <think> blocks from local reasoning models. OpenRouter, Gemini, and local endpoints all speak this one protocol.

Roadmap

• Polish pass: shortcut-collision check, options UX, a "(partial)" note when a very long transcript gets sampled down.
• Package for the Chrome Web Store.

Credits

• YouTube caption retrieval is adapted from the technique in the open-source youtube-transcript package (MIT), the InnerTube ANDROID-client recipe, reworked here for a browser content-script context.

License

MIT © 2026 Marc Tooze. You bring your own LLM endpoint and API key; TD;DW never ships or proxies one.

Not affiliated with, endorsed by, or connected to YouTube, Netflix, Prime Video, Disney+, Jellyfin, or any LLM provider. All trademarks belong to their respective owners.