Building a personal knowledge base from short video

Two decades of note-taking software trained an entire generation of expectations. Full-text search that returns in milliseconds. Links between documents, and backlinks showing what points here. Tags, filters, saved queries, a graph view if you want one. Whatever you think of the more elaborate systems, the baseline they established is real: text you wrote down is text you can get back.

Short video arrived with none of that and inherited none of it. The result is a familiar split personality — the same person maintains a tidy, searchable, cross-linked notes vault and a saved folder with two thousand videos in it and no way in. It is not that they gave up on organising the second one. It is that everything they learned about organising the first one has no purchase on it.

Three properties that make a pile into a knowledge base

It is worth being concrete about what the notes tools actually provide, because "organized" is too vague to build against. Strip the feature lists back and three properties do nearly all the work.

The first is full-text search over the content itself, not over labels attached to it. The crucial word is content: you can find a document by a phrase buried in its fourth paragraph, one you never anticipated needing and never tagged. That is what makes a collection tolerant of growth, because you do not have to predict your future questions at filing time.

The second is a stable address for every item — something you can link to, return to, and cite, which still resolves to the same thing next year. Addressability is what allows items to reference each other, and a collection whose parts can point at each other is worth considerably more than the same parts in isolation.

The third is classification that does not depend on you filing things by hand. Any scheme requiring a decision per item degrades in exact proportion to how busy you are, and you acquire the most items when you are busiest. Classification has to be derived from the item, or it has to be so coarse that the decision costs nothing.

Text gets all three almost for free, which is why nobody notices they are the requirements. Text is searchable because it is already the thing search operates on; a note has a filename and a path; and classification can be inferred from the words on the page. Video has exactly none of these by default, which is the whole problem in one line.

Video only gets these through derivation

You cannot search a video. It is worth resisting the loose phrasing here, because the loose phrasing is what makes the rest of this confusing. What actually happens is that text is extracted from a video — spoken words, on-screen text, the caption — and then you search that text. The video is never queried. A shadow of it is.

Once that is clear, several otherwise mysterious behaviours become predictable. The quality of your library is bounded by the quality of the extraction, not by the interface in front of it. Two products with identical grids, identical filters and identical search boxes will differ enormously in whether they find things, and the difference lives entirely in a pipeline you never see: how the audio was transcribed, which frames were selected before any text was read off them, and how the resulting fragments were assembled into something worth indexing.

This is why the extraction pipeline is the product and the folder view is decoration. It is an uncomfortable thing to be true, because the pipeline is invisible and expensive and the folder view is what gets screenshotted. But a beautiful library over thin derived text is a beautiful library you cannot search, and users experience that as "it never finds anything" without ever suspecting where the failure occurred.

Addressability, at least, comes free — a reel already has a canonical URL, and a library that stores that link inherits a stable address without doing any work. Classification sits in between: derivable from the extracted text, and only as good as that text, which is why a confidence score attached to a category is more honest than a bare label.

Retrieval beats collection

The other half of this is behavioural, and no amount of machinery substitutes for it. A knowledge base earns the name through use; the number of items in it is not a measure of anything. A few habits carry most of the difference between a library you consult and one you merely maintain:

• Review a slice, never the whole library. Take one category at a time and read it as a list rather than watching anything. The purpose is not to enjoy the items again — it is to notice which ones you no longer recognise from their text alone, because those are exactly the ones a future search will never bring back.
• Delete on sight. Every item kept out of vague sentiment is one more candidate competing with the thing you were actually looking for. Removing it is not a loss of information; it is a small permanent improvement to every search you run afterwards.
• Do not build a tag hierarchy. It needs constant tending to stay coherent, it splits the moment two branches could both apply, and it stops being trustworthy the week you get busy. A flat handful of distinctive words does the same retrieval work and asks nothing of you.
• Accept that most of what you saved was entertainment. That is not a failure of curation, it is an accurate description of short video. A knowledge base that pretends otherwise is mostly padding.

What short video is genuinely good for

None of this effort is worth it unless the format carries something text cannot, and it does. Short video is the best medium there has ever been for demonstration — for showing a physical process at the speed it actually happens, from an angle that makes it legible.

How much pressure to apply to a sharpening stone. What a properly kneaded dough looks like when it is ready and how it behaves differently thirty seconds earlier. The exact wrist rotation in a knot. The sound a bearing makes when it has gone. Written instructions for any of these are notoriously bad, not because writers are careless but because the information is genuinely non-verbal — it is a rate, a texture, a feel. Thirty seconds of footage transmits it instantly and a page of prose does not transmit it at all.

Naming that honestly also names the inverse. Video is a poor container for anything you will need to quote, cite, compare side by side, or scan for one number. A reel listing seven statistics is worse than a paragraph listing seven statistics in every way that matters after the first viewing: you cannot skim it, you cannot copy from it, and checking one figure means watching until it appears. If what you are saving is essentially text that happens to be narrated over a background, the video form is pure overhead and a screenshot would have served you better.

Where it breaks

Two categories of content defeat derivation outright. A silent, purely aesthetic clip produces no transcript and often no on-screen text, so there is nothing to index and no query will reach it. Content whose meaning lives in a fast-moving edit — where the substance is the juxtaposition of shots rather than anything said or written — is similar: the words that come out are a poor proxy for what made it worth keeping. Both are findable only by browsing, and it is better to expect that than to discover it while hunting.

The more serious limit is structural. A derived library is an index of things that live somewhere else. When a creator deletes a post, the link dies and what remains is your record of a video you can no longer watch — the transcript, the extracted text, the category, all describing something gone. There is nothing to be done about that short of re-hosting other people’s work, which is not a thing an organizing tool should do. Index, not archive — a distinction that costs nothing to hold and shows up on exactly one day, which is the day you finally go looking.

It is tempting to file all of this under tidiness, which makes it sound optional. It is closer to an access problem, and the stakes are quietly rising: a growing share of practical know-how — the kind that used to arrive through manuals, evening classes, or someone standing next to you who would show you once — now reaches people first as a two-minute clip from a stranger. That is a genuinely good development. It is also the only substantial body of knowledge most of us hold that has no index at all, and building one is the difference between having watched something and being able to use it.