Turning a YouTube video into notes you can actually verify
Seven things happen between pasting a link and getting your cards — fetching captions, aligning timestamps, breaking lines, translating, reviewing, building cards, anchoring them. Here is what each step does, and what to do when one of them fails.
You have probably been here: a 40-minute interview in a language you read slowly, a headline that hooks you, and no 40 minutes to spare. You turn on auto-captions, get a wall of text, and give up two minutes in.
Seven things happen when PandaRead processes that video. Knowing what each one does means you can tell where things went wrong instead of just retrying and hoping.
Where the subtitles come from
Subtitles are not simply there for the taking. YouTube blocks requests coming from data-center IPs, so our servers usually fail to fetch them. That is why the captions are pulled inside your own browser — and the main reason the extension exists. It borrows your ordinary session with YouTube, catches the caption data the player itself receives, and hands it to the pipeline.
If the video has no captions at all, we fall back to speech recognition. The two paths differ in precision:
- Platform captions are usually sentence-level, and human-stenographed tracks lag behind the audio by half a line to begin with;
- Speech recognition gives us word-level timestamps — every word carries its own time.
Only the latter produces subtitles that sit tightly on the audio. So occasionally a video with official captions ends up less precisely synced than one without. That is the source material, not us cutting corners.
Why line breaks are harder than translation
Splitting continuous speech into subtitle lines looks trivial. It is the most error-prone step in the pipeline.
The reason: each segment is translated independently. Break in the wrong place and no amount of translation quality saves you, because the break is inherited by the translated line verbatim.
Cut an English sentence after “and” and the Chinese line ends on a dangling connective. Netflix's own subtitle rules forbid conjunctions, prepositions and articles at the end of a line.
So our rule is: sentence-final punctuation wins, and before any hard cut we walk backwards to avoid a forbidden ending. Each of the 11 source languages has its own list of words that may not end a line. Chinese is judged by the final character rather than the final word, because that is where a Chinese line actually breaks badly.
Translation is followed by a review pass
A literal pass is only the draft. The review pass handles localization: forms of address, idioms and slang, units and proper nouns. Japanese 先輩 should not become "senior schoolmate", and "ballpark figure" should not turn into a sentence about baseball.
That pass has one hard rule: never change facts, numbers, or references. Making a line read better is fine. Changing what it means is not — that is the line between a translation tool and a rewrite.
Cards are curation, not extraction
Only once the full transcript exists do we build knowledge cards. The principle is few and precise: 6–8 for thin material, 10–15 for dense material, never more than 15.
- Every card must carry visual load — a wall of text earns no second look;
- Structural molds (quadrants, funnels, timelines) are used only when the source genuinely has that shape; we do not invent one;
- Every card is anchored — one click jumps back to that second of the video.
That last point matters most. A breakdown can be wrong, but it must be verifiable: whatever a card claims, you can always jump back to the source and judge for yourself whether we read it correctly. A summary you cannot trace back is asking for unconditional trust.
When something goes wrong
| What you see | Which step | What to do |
|---|---|---|
| "The server can't fetch subtitles" | Fetching | Install the extension and parse from the video page |
| Subtitles lag half a line behind the audio | Fetching (stenographed track) | Compare against a video with auto-generated captions |
| Lines break in strange places | Line breaking | Send us the link and a timestamp |
| A section is clearly missing from the cards | Cards | Hit "Regenerate" — it reruns cards only, not the translation |
One last thing: parsed videos are cached. Re-parsing the same link hits the cache instead of rerunning the pipeline, so to see the effect of a fix you need a video that has not been parsed before.