I am a partner with a CLE organization called CLEWebinars. I record the webinar in advance and send the recording over. The folks at CLEWebinars do an amazing job of editing my recordings. Once they have completed the editing, they make them available both OnDemand and through a Simullive replay. During the Simullive replays, I am available to answer questions via chat.
When I record the webinars, I frequently have starts and stops so I am never sure exactly how long the final result will be until they are edited. Sometimes I record straight through and have very few interruptions, but that is rare. Most of the time, I have a coughing fit, or my dogs decide it is time to wrestle on my lap, or I don’t like how I phrased something, so I pause, count down, and resume the recording, telling the editor where they should stop and start.
For my updated Bad Review? Bad Response? Bad Idea. webinar, I went over by 15 minutes. I am not particularly surprised; it felt long when I recorded the program. There were a lot of updates since I last recorded this program in 2024. Especially as it relates to artificial intelligence. As a result, today, after he had completed the edits, the editor reached out and asked me if I could tell him what could safely be cut out.
Claude Helped Me Decide What to Cut
Did I want to watch my own video and figure out what should be sacrificed? I did not. So, of course, I turned to Claude.
This was my process:
- Download the video
- Put it into a folder to which I gave Claude Cowork access
- Explained the situation and asked for recommended cuts in a Word document, explaining each cut
- The final document noted the times for the editor, though it did warn me that the times could be a bit off. The document also made it very clear exactly what cuts it was recommending and why.
Claude quickly gave me a document with a chart. The chart identified the time, the length, what it was, and why it could go. The document also explained what it left alone and, because I tell it to do so in my standing orders, any inferences it made when choosing the cuts.
Here are some of the things Claude found:
- Tool-by-tool detail. Claude noted rightly that this is the part of the lecture that was not going to remain pertinent for long, since technology changes so rapidly.
- Dead air, not much, but a bit here and there that could be cut down. This alone helped with several minutes.
- A tangent that was relevant, but not necessary.
- Sometimes, to emphasize something, I will restate it more than once. Claude recommended removing some of those.
- Claude even told me that it had some concern about cutting one thing, a recitation of a case, but it noted that I had several other similar cases, and the case it wanted to cut was the shortest.
The end result meant that instead of sitting and listening to an hour and fifteen-minute recording, I was able to send a list of timed cuts, with explanations, to the editor so he could get the work done quickly, before this Friday when my updated webinar premieres.
Workflow for Editing Videos
Claude Fable 5, working in Cowork, is capable of taking a video from its raw form, recommending edits, editing it, and minimizing the work necessary by CLEWebinars. I have doubts that Opus 4.8 has the judgment that Fable has, as far as making edits that are not obvious.
In addition, Claude warned me that its timestamps could be off by a second either way. This is a serious limitation in terms of having Claude make the edits itself, but not a particularly big problem if a human is making the edits. I know many people use AI to edit videos, but you will want to keep this limitation in mind if you are handing a completed video over to AI to have it make cuts for you.
This is the workflow I recommend for editing videos that are already recorded:
- Ask the tool to recommend edits and to give you a timed sheet of what it would edit and why
- Review the edits and make sure they are what you want
- Ask the AI to actually edit the video
- Review to make sure that the edits are accurate
The first two steps and the final step are the verification process that I always build in when I am working with generative artificial intelligence. I build them in not only because I am an ethics lawyer, but because my frequent use of GAI has shown me repeatedly that the tendency of GAI to both hallucinate, and to use a human word, disobey, could cause it to make cuts I do not want in a way that could make the result inaccurate. Or, as Claude warned me, it might have the timing off at the start or finish of a cut, which would be a problem.
AI is Developing Rapidly
The amazing thing about AI is that for the most part, its only limits are what you can think of with your mind. Yes, there are certain technical limitations, but every month more and more of those limitations disappear. It wasn’t all that long ago when AI told me it couldn’t “look” at or edit videos. Now it can do both.
It is common for me to find that something AI couldn’t do in May it can do in July. Not because anyone told me that a change came, but because I periodically ask about things I previously learned AI could not do, or because the AI suggests it can do something for me. It never hurts to ask a second time, especially if a new model is introduced.