The useful part of the recording is somewhere after the introduction. Someone demonstrates the difficult step, a question changes the explanation, and a correction arrives ten minutes later. A summary can flatten those moments into a tidy paragraph. An editor still needs to find them.
Make a scene map instead: a short list of moments with time ranges, a description of what happens, and a note about why a reader or editor might care. The aim is to make the recording easier to revisit before deciding whether to cut, quote, or publish anything.
This workshop combines a transcript, a few reference images, and an AI-assisted first pass. It is designed for a recording you own or are permitted to process, such as your own tutorial. Do not begin with private footage whose participants or owner have not agreed to this use.
Gather three views of the same recording
Keep the original video unchanged. Make a working copy and note its duration. Use one player as the timing reference so the transcript and image samples can be compared with the same timeline.
Obtain a transcript with timestamps from an approved tool. If using Whisper, read its model card: OpenAI documents that it can generate words absent from the audio and that performance varies across languages and accents. A transcript is a finding aid, not a verified quotation.
Next, capture reference frames at a regular interval and at obvious visual changes. FFmpeg's filter documentation describes frame-rate conversion and frame selection; a video editor can also export stills manually. Use file names or visible labels that preserve the approximate source time. A contact sheet gives the reviewer a quick view of slides, demonstrations, and changes of setting.
Keep the interval modest for the first pass. Sparse samples will miss actions between frames. Dense samples create more material to inspect. Neither is a replacement for watching the candidate moments.
Ask for moments, not a finished highlight reel
Give the assistant the timestamped transcript and permitted reference frames. Describe the audience and the question the video answers. Ask it to propose a small number of candidate moments, preserving supplied times and labeling uncertainty.
For an invented repair tutorial, the map might contain the setup, the close-up that shows a part's orientation, a common mistake, and the final working result. These are examples of useful scene types, not claims about an actual recording.
Request a plain table with four columns: approximate time, what is said, what is visible, and why the moment may be useful. A row should distinguish a speaker's explanation from something shown on screen. If the assistant received only transcript text, it should not describe an unseen action.
Also ask it to flag corrections and qualifications. The most quotable sentence may be followed by a limitation that changes its meaning. A good scene map helps the editor find that context rather than making the first interesting phrase easier to detach.
Watch before selecting
Open each candidate in the original player. Start before the proposed beginning and continue past its end. Confirm the words, the action, and the time range. Correct the map as you go.
Listen for a name or number that the transcript may have misheard. Watch for a visual demonstration that the transcript barely captures. If a later correction changes the passage, connect the two moments in the map.
Keep a note of why you rejected a candidate. Perhaps the scene needs material outside the proposed clip, the audio is unclear, or the image reveals information unsuitable for reuse. Those reasons can prevent the next editor from rediscovering the same problem.
At this point, the model has helped you search. A person has made the selection. No clip has earned publication approval merely by appearing in the table.
Make the map useful to someone else
Save a short introduction identifying the recording, duration, audience, and version. Follow it with the checked moments and links or time ranges that a colleague can open in the original. Include the later qualification beside the earlier claim where necessary.
Keep quotations marked as checked or unchecked. Label the contact sheet as a navigation aid, and retain the video as the source. If the final edit will use a different export, check that the times still match; an added opening can shift every reference.
Then hand the map to the person making the edit. They should be able to find the demonstration without reading a full transcript or trusting a generated description of a frame.
This process does not determine reuse rights, replace consent, or certify captions. It produces something smaller and immediately useful: a route back to the recording. The useful minute stays attached to the explanation that made it useful in the first place.
Sources and limitations
- OpenAI, Whisper model card: checked October 15, 2026. Supports intended transcription use, hallucinated-text limitations, language/accent variation, and the need for context-specific evaluation. No Whisper run was performed for this article.
- FFmpeg, Filters Documentation: checked October 15, 2026. Primary documentation for frame-rate conversion and available video filters. The workshop's interval choice, table, and editing steps are editorial proposals, not documented accuracy guarantees.