Video clipping from session recordings without burning a week of editor time

The show closes on a Thursday. By Monday, 42 session recordings are sitting in a shared drive, someone from marketing has asked for "a few clips for social", and the editor who has to make them opens the first file and starts watching. Video clipping from session recordings looks like an editing job. It is a reading job with an editing job attached, and the reading is where the week goes.

Forty two sessions at an average of 52 minutes is 2,184 minutes of footage, which is 36.4 hours. That is the cost of watching everything once, at normal speed, with no cutting, no rendering and no revisions. Nobody budgets for it because nobody puts it on the estimate, and then the clips arrive three weeks late and the moment has passed.

Why selection costs more than editing

Finishing a clip is a known quantity. Trim the head and tail, add a caption track, set the title card, export. An experienced editor working from an approved timestamp does it in twenty to thirty minutes, and the variance is small because the work is the same every time.

Deciding which 90 seconds out of 52 minutes is worth cutting has no ceiling at all. It is a judgement call made against an unstated standard, which means two editors produce different clips from the same session and neither can say why the other is wrong. So the work expands. The editor watches the session again to check whether the better moment was at minute 34. Somebody in marketing asks for something punchier and the search starts over.

The fix is to move the judgement out of the timeline and into a written rule that a person who is not an editor can apply.

A selection rule you can hand to somebody else

Here is the one I would start with. A passage qualifies as a clip if, inside 90 seconds, it contains at least one of three things: a claim the speaker is willing to own, a number with its unit and period attached, or a disagreement between two people on the panel.

Everything else is context. A speaker introducing themselves is context. A moderator setting up the question is context. A speaker agreeing with the previous speaker is context. None of it survives being lifted out of the session, because the thing that made it work was the fifteen minutes around it.

The rule is deliberately narrow, and narrow is the point. Applied to a transcript, it produces a small number of candidate windows per session with start and end timestamps, and a person can apply it while reading rather than while watching. A 52 minute session transcribes to roughly 8,000 words. Scanning that against three tests takes about twenty minutes.

Run that over the same 42 sessions and selection costs 840 minutes, which is 14 hours. Against the 36.4 hours of watching everything, you have taken about 22 hours out of the job before an editor has touched anything.

What does clipping actually cost per clip?

Put the numbers together on one edition and the figure stops being a mystery.

Selection on transcripts: 14 hours. Say the rule surfaces an average of two qualifying windows per session, which is 84 candidate clips. Finishing at 25 minutes each is 2,100 minutes, or 35 hours. Total editor time is 49 hours. At an illustrative 55 per hour, the programme costs 2,695.

Divide that by the 84 clips made and the unit cost is 32.08. Now divide by the clips that actually got published, which in most first attempts is well short of the clips made. If 60 of the 84 go out, cost per published clip is 2,695 divided by 60, or 44.92. Clips per editor hour across the whole job is 84 over 49, which is 1.71.

Those three figures are worth putting in the same place every edition. The gap between 32.08 and 44.92 is the money spent on clips that nobody used, and it is the fastest thing to fix, because it is a selection problem and selection is now written down.

One caution about the denominator. Published means it went out on a channel you own or paid for, with a date. It does not mean the file was delivered to a folder and approved in principle. Teams that count delivered clips instead of published ones get a flattering unit cost and lose the only signal that tells them the rule needs tightening. If 24 clips sat in the folder unused, somebody chose not to use them, and the reason is usually that they were selected against a rule the person posting them does not share.

Track the same figures by session type as well as in total. Keynotes and panels usually clear the rule several times over. Sponsored sessions and product theatre often clear it never, and knowing that in advance saves you the reading pass on twelve files next edition.

Who should actually do the selection

Once the rule is written, the reading pass does not need an editor. It needs somebody who understands the market well enough to tell a real claim from a pleasantry, which describes the conference producer who booked the speakers, the editor of the show's media title, or the analyst who wrote the audience brief. Any of them can read a transcript and mark timestamps.

That matters for cost and it matters more for queue length. An editor is a scarce resource with a rendering machine attached, and the fortnight after a show is when everything else also wants that person. Selection done by somebody else in parallel means the editor starts the finishing work on day two with 84 timestamps in hand instead of starting the reading on day two and the cutting on day nine.

It also gives you a review that is worth having. The person who marked the windows and the editor who cut them disagree occasionally, and those disagreements are the only evidence you will get about whether the rule is set correctly. Log them. If the editor is rejecting one candidate in three, the rule is too loose. If they are asking for windows the rule never surfaced, it is too tight.

Batch the work session by session

The common failure after the rule is in place is scheduling. A request arrives for three clips for the keynote, then two for the sponsor session, then one more for the keynote because the first three were wrong. Each of those is a separate file open, a separate project set up, a separate export queue.

Work one session at a time, all the way through, and the setup cost is paid once. Pull the transcript, mark every qualifying window, cut all of them, export all of them, then close the session and never open it again. The editor keeps the speaker's voice and cadence in their head across all the clips from that session, which is worth something on caption accuracy alone.

The corollary is that you decide the clip count for a session before you start it, from the rule, and you do not go back. A session that yields one clip yields one clip.

What the markup wants back from selection

There is a downstream reason to record the start time at selection rather than reconstructing it later. Google's video structured data documentation lists three required properties for a Clip: a name, which is a descriptive title for the content of the clip, a startOffset expressed as the number of seconds from the beginning of the work, and a url pointing at the start time. The same documentation sets a minimum video length of 30 seconds and requires that no two clips defined on the same page share a start time.

Every one of those is free at selection time and expensive afterwards. The person reading the transcript already knows the start second and already has to write a descriptive line to explain why the window qualified. Capture both in the selection sheet and the markup is a data entry job later. How that markup is built and tested is a separate piece of work on video structured data.

Does short actually work better?

The 90 second ceiling deserves a check rather than an assumption. Wistia's 2026 State of Video Report, built from more than 900 professionals surveyed plus over 13 million videos and 79 million hours of viewing data, reports that the shorter the video, the higher the engagement rate. The same report makes the honest counterpoint in its own words: "a 10-minute video with a lower engagement rate still gets more total watch time than a 1-minute video with a higher engagement rate."

So the rule is about cost and consistency more than about reach. A 90 second ceiling makes selection decidable and keeps the finishing time predictable. If your own numbers show two minute clips holding attention on the platforms you post to, move the ceiling and keep the rule.

Where this stops

The transcript is a lossy record of the session, and the rule inherits every gap in it. A product demo where the speaker says "and you can see here" for forty seconds reads as nothing on the page and plays as the best moment in the session. An audience question that lands because of the pause before it looks like filler in text. A speaker whose delivery carries the point will be systematically under-selected against a speaker who writes in complete sentences.

There is no clean fix. What works is a small override budget: whoever ran the room flags up to three sessions where the recording is better than the transcript suggests, and those get watched. Three overrides on 42 sessions is two or three hours, which the transcript pass has already paid for.

The other limit is that none of this tells you whether you are allowed to publish the clip. Selection produces a list of moments, and what the speaker actually signed decides which of them can leave the building.

Take one session from your last show, pull the transcript, and mark every window where a claim, a number or a disagreement fits inside 90 seconds. Count them. That count times the number of sessions is the size of your clip programme, and it is the first honest input into the wider question of what a recording library is worth and into how much editorial capacity the content operation needs next year.