From 400 clips to 11 selects

Saturday came back as 403 clips. The cut needs eleven of them. The eleven already exist, sitting in the grid with the other 392, and the entire job is making those 392 step aside. Scrubbing is the slow way, and it stops scaling at about fifty clips; memory is worse. Filtering is the fast way, and filtering only works on footage somebody logged. That is the honest trade: minutes of logging on the front buys back hours of hunting on the back, every single time you return to the material.

Every confirmed value is a way back in

The mechanic is simple: any value you confirm becomes a facet you can filter the grid by. Ratings, flags, colors, keywords. Technical facts like codec, duration, frame rate. Evidence like has-transcript, people-in-frame count, on-screen text. And your own fields: action, moment, athlete, whatever your palette defines. Each filter is a chip; chips stack; the grid shrinks. The pile stops being a pile the moment it can answer questions, and every question you answered once (who is this, was it good) is a question you never answer again. One design choice does the load-bearing work here: pending AI suggestions are deliberately not facets. Until a human confirms a value, it cannot appear in a filter, an export, or a filename. The filter panel speaks only in decisions, which is what makes its answers trustworthy: when you filter to flagged, you get takes somebody actually vouched for, not takes a model liked.

Three fields carry a whole day

This part works in any tool with keywords and ratings, no install required. Ask three questions of every clip: who is in it, is the take usable, what kind of moment is it. Closed answers, decided before the shoot: if the list says TD, then touchdown is not allowed to exist, because forked vocabulary is how archives quietly die. Log at offload while the day is fresh, not at wrap. Three fields sounds thin until you do the arithmetic: three confirmed values per clip across 400 clips is 1,200 hooks back into the footage, and any two of them intersected will collapse the day to a handful. The failure mode is the opposite instinct, the twenty-field template that anticipates every future need and gets abandoned by clip thirty. A three-field log completed beats a twenty-field log imagined. You can always add a field next shoot; you cannot retroactively log Saturday.

Stack until the grid gets small

Here is Saturday collapsing in ClipLogger. Filter to the athlete: the project's cast resolves him whether the evidence in a given clip was his face group or his jersey number, and 403 becomes 61. Add the flag, takes somebody marked good on the day: 61 becomes 19. Sort what is left by the quality score, the on-device aesthetics pass that ranks the technically clean takes above the soft ones: the eleven float to the top of nineteen, and the whole pass took under a minute. The search box runs the same play in text for people who think in queries: action:td flag:good has:transcript, plus free text that matches names, notes, transcripts, and on-screen text, so "the interview where he mentions the transfer" is a query and not an afternoon. Neither route is clever. Both are intersection, applied to values that exist because ten seconds of logging happened at offload.

Sort with the model, decide with the flag

The panel keeps the model's opinions and yours distinct, and the distinction is visible. Log status is a tri-state facet, Unlogged, AI-suggested, Confirmed, so "everything still waiting on review" is one click, and so is "everything no model has touched." Each AI value carries provenance, which engine produced it, local or cloud, shown on the clip. I use the model's lenses happily: quality sort surfaces the sharp takes, people-in-frame counts find the crowd shots, cull candidates sweep the junk the other way. But a lens is not a verdict. The flag is the verdict, and only verdicts reach exports, renames, and the filenames an editor will trust. Search by what the model found when you are exploring; search by what you decided when it counts. The day you conflate the two is the day your selects include a beautifully exposed shot of the wrong kid.

Save the query, not the pile

When a filter stack earns its keep, keep it as a smart bin: a saved query that re-evaluates live as new clips get logged, with pins for the exceptions that belong despite the query and exclusions for the ones that match but do not. Bins are query views, so nothing moves and nothing copies; "Week 4 selects" can exist forever without duplicating a frame. Selects folders rot the moment someone forgets to maintain them. Queries stay current because they are re-asked, not remembered. And because every confirmed value rides the sidecar beside the file, the same filters work next season, on another machine, when Saturday is forty weekends old.

The eleven were in the grid all along; logging is what makes the other 392 step aside.