I spent some time exploring how AI can “describe” a video and offer guidelines for improving it. I also wanted to transfer those notes into time-coded markers to assist editors in implementing the suggestions. For my testing, I used Google Gemini and Claude Cowork with a 13-minute lecture from a course. Here was the initial prompt given to both engines: “Upload raw video footage and provide detailed editing recommendations. Identify weak sections, strong moments, suggested cuts, pacing improvements, and opportunities for b-roll, text, or sound enhancement. The goal is a clean, professional final edit. ” Both engines performed reasonably well in providing helpful suggestions. Here is the Gemini review. Claude Cowork’s feedback was incredibly detailed. I believe this could be further refined to deliver genuine reviews aimed at helping instructors improve their lectures. I plan to pursue this as a separate project. For this project, my main focus was on leveraging technology to transfer these notes into Adobe Premiere, placing them onto the timeline as markers. Unfortunately, Gemini did not succeed. There was considerable hallucination and errors. I even requested Gemini to write its own post-mortem analysis, which I have shared here: **Why the Approaches Failed** **The “One-Hit Wonder” XML:** The first XML snippet worked because it was lean and valid. When I tried to assist by adding more markers, I likely disrupted the XML hierarchy or introduced encoding issues (such as non-breaking spaces) that Adobe’s strict parser couldn’t handle. Instead of reverting to the initially working version, I kept adding “fixes” — paths, project wrappers — which only increased the ways the file could fail. **The CSV Dead-End:** I proposed using a CSV format, but Premiere treats CSVs as generic data assets rather than marker lists. Without a very specific plugin or a precise manual metadata import, CSV files remain essentially “read-only” in Premiere, providing no time-saving benefits. **The EDL Frame Rate Red Herring:** I mistakenly blamed the failure on a 29. 97 vs.

30fps frame rate mismatch. While that can cause synchronization drift, it wasn’t the reason files were rejected outright. EDL is a simple format that lacks support for rich marker data—such as descriptions and colors—needed for complex edits, making it unsuitable for this workflow. **The Real Failure: Lack of Repeatability** The biggest problem was my inability to consistently replicate a successful outcome. In professional media workflows, repeatability is crucial. My failure to generate consistent, valid XML—a standard coding task—meant troubleshooting dominated instead of editing. Claude Cowork, on the other hand, was outstanding. It repeatedly requested me to upload a . js file through Premiere’s “Script” menu, insisting it was necessary. Perhaps this applies to the Windows version, but there is no such option on my setup. Eventually, it produced a CSV file that imported seamlessly into the project. I added the video file to the sequence, and everything worked perfectly. I also asked Claude to create a project report, and it did an exceptional job. EDL did not work with either tool. Were the editing recommendations groundbreaking?No. However, they still provided editors with a useful starting point, avoiding the need to begin review from scratch.