# How to plan a compilation video from old uploads > A compilation is a selection problem before it is an editing problem, so plan it backwards. Canonical URL: https://vivu.ai/guide/how-to-plan-a-compilation-video-from-old-uploads A compilation is a selection problem before it is an editing problem, so plan it backwards. Define the theme narrowly enough that you can hold a clip up against it and say yes or no. Build a candidate list of moments with timecodes and the source file each one lives in. Cut that list down harder than feels comfortable. Only then open the editor. Compilations that drag are almost always compilations where the selection step was skipped and the theme was decided while cutting. ## Make the theme testable "Best moments" is not a theme, because every clip passes. A theme you can test looks like every time a guest admitted they were nervous, or every shot where the product appears on a phone screen, or every question that got the same wrong answer. The test matters more than the wording: you should be able to look at a clip for five seconds and know whether it is in. This also decides your runtime later, because a tight theme with twelve real instances is a better video than a loose one with forty. Write the test down before you start looking. The failure mode is drifting the theme to fit whatever you happen to find, which produces a video that is about nothing in particular. ## Where candidates come from Your own memory of the obvious ones, first, because those come free and they anchor the structure. Then whatever notes, chapter markers, or descriptions you wrote at the time. Comments on old uploads are underrated here, since viewers timestamp the moments that landed. Transcripts help when the theme is something people said and you remember the phrasing they used, though they are blind to anything visual and to anything phrased differently than you expect. [Working from a transcript](https://vivu.ai/guide/how-do-i-use-a-transcript-to-find-good-video) has its own workflow and its own ceiling. The brute-force option is rewatching everything at double speed with a notepad open. For a small back catalogue this is genuinely fine and often faster than setting up anything smarter. ## Two passes, not one Whatever method you use, separate the question "which videos might contain this" from "where exactly in them". A quick pass that returns whole videos as candidates is useful for narrowing the field, and it is all you need if your uploads are short, but it leaves you dragging the playhead to find the actual seconds. A thorough pass that returns specific ranges takes longer and is what you want once the field is narrow. Doing the fast pass first and the precise pass second on the survivors is simply cheaper than being thorough about everything, and it keeps the list moving. A search layer does this shape of work: uploads go into a project, get indexed once, and then a themed question returns time ranges from across different videos, each with a line explaining why it matched. [Vivu](https://vivu.ai/platform) is one of these, and for a compilation the relevant part is that one query gathers scattered instances into a single list instead of you opening each old upload in turn. What it will not do is decide the theme for you or cut anything together. ## Expect borderline clips, and decide in advance Some candidates will be adjacent to the theme without being on it. A clip where someone says they spent years learning to relax on stage is related to a compilation about nerves, and whether it belongs is a judgment call about pacing and about what the video is arguing. No search method resolves this, and the useful preparation is deciding your bias beforehand. If the compilation is a highlight reel, exclude the indirect ones. If it is making a point, one or two indirect clips can be what carries it. ## Get the source files back Plan the edit around the files you still have, not the videos that are still online. Published uploads have been re-encoded, and they usually carry intros, end screens, and licensed music you cannot cleanly separate. If your masters are gone for the older material, you are working from the published render and your quality ceiling is set there, which is worth knowing before you promise yourself a 4K compilation. Rights are a real constraint too, since anything in the original that was licensed for that video is not automatically licensed for a new one. ## When the planning step is not your problem If the compilation draws on your last handful of videos and you already know the six moments, there is nothing to plan. Write them down, pull the files, cut. The selection step earns attention when the catalogue is large enough that you cannot hold it in your head, or when the theme is one you never thought about while making the originals, which is the case where your notes are guaranteed not to help. [Hunting inside one long recording](https://vivu.ai/guide/how-do-i-find-viral-moments-in-long-videos) is a different job with different tools. The way to tell which side you are on is to try listing candidates from memory for ten minutes. If you get a full list, your archive is small enough and the rest of this is overhead. If you get four clips and a strong feeling that there are more, the gap between those two numbers is the whole problem, and it is a retrieval problem rather than an editing one. ## FAQ ### Do I need the original files or can I use my published uploads? Use the originals if you still have them. Published uploads are re-encoded for streaming, which means recompressing them into a new edit stacks artifacts, and they typically include intros, end cards, and burned-in graphics you would have to crop or cut around. Downloading your own uploads as a source is a fallback, not a plan. If you are shooting new material now, this is an argument for keeping masters on cheap archive storage rather than deleting them once a video is live. ### What if I deleted the source footage for older videos? You work from the published render and accept the ceiling that sets. In practice this is survivable for a compilation, since the clips are short, mixed with newer material, and often watched on a phone. The two things to watch are music you no longer have rights to use in a new context and any on-screen branding from the original video that will look odd in the compilation. Both are easier to spot during selection than after the edit is locked. ### Can an AI pick the clips for a compilation? A search can return candidates that match a description, which is most of the tedious part. Deciding which candidates actually belong is not something to hand off, because the criteria are about pacing, about how much the audience already knows, and about what the video is trying to say. Expect to discard a real share of anything a search returns, including matches that are technically correct and dramatically flat. ### How do I stop a compilation from feeling like a list? Order the clips by shape rather than by date. Chronological order feels natural to make and usually plays as a list, because nothing escalates. Putting your second-strongest clip first, the weaker ones in the middle, and the strongest at the end gives the video an arc even when the clips are unrelated. Cutting the runtime by a third after the first assembly tends to help more than any transition.