He was writing a video about markets, and needed a shot of a market at night. He was certain he had filmed one on a trip two years earlier: a fruit cart, a hanging lamp, a man stacking crates. He opened the external drive and found eight folders named after cities, and inside each one, hundreds of files called C0074, C0075 and the like.
After two hours of opening and closing he shut the laptop and bought a clip from a paid library. The shot he had filmed himself was two steps away from him, and he could not reach it.
This is not a storage problem or a drives problem. It is an indexing problem: your material exists and cannot be searched. AI solves exactly half of that problem, and leaves the other half to you.
Two halves: what you name, and what the tool describes
The first half: structure and names. This is your work and you do not delegate it, and it takes five minutes at the end of every shoot day. One folder per shoot, named starting with the date, then the place, then the subject: 2026-03-12_aswan_night-market. Date first so it sorts itself, and words joined by a single separator rather than spaces.
The second half: describing the content. This is what no human can do for hundreds of files, and it is what you delegate. Current tools can read a clip and write what is in it, transcribe every spoken word in an interview, and tag the location and the shot type.
The distinction matters: the first half tells you which folder to search, the second tells you which file. Without the first, the second becomes a heap of descriptions with no context.
How to build the description layer in practice
The order that works:
- Transcribe the speech first. Every interview and every clip with audio becomes timed text. This is the single biggest gain the tool gives you, because most of what you search for in interviews is a sentence rather than an image.
- Ask for a short visual description per clip. One line: what is in the frame, the shot type, the time of day. One searchable line beats an eloquent paragraph.
- Collect everything in one file. A text file or a sheet in the shoot folder, holding the file name, the description and the transcript. Searching one file is faster than any tool.
- Add what the tool cannot see. These are the fields that save you from real trouble.
A six-field card per shoot
| The field | What goes in it | Who writes it |
|---|---|---|
| Date and place | The shoot day, the city, and the specific location | You |
| People | The name of everyone who appears, and in what capacity | You |
| Consent status | Signed, verbal, or none, and what must not be published | You |
| Visual description | What is in the shot and its type | The tool |
| Spoken text | The speech with its timings | The tool |
| Quality | Usable, or shaky, or out of focus, or bad audio | You |
The consent field is the most important one in the whole table, and the tool knows nothing about it. Knowing that a clip has no signed release before it goes into a new film prevents a problem that cannot be fixed after publication.
The quality field saves serious time on every future project: a clip annotated "shaky in the last third" never gets opened twice.
What you never delegate to the tool
- Consent and permission status. A legal and ethical fact whose source is your memory and your paperwork, not the image.
- People's names. The tool writes "a man in his fifties"; you know he is Mahmoud, who owns the workshop.
- Judging what is any good. The tool describes; it has no idea which shot has life in it.
- Checking what it says. Transcripts get names, numbers and local terms wrong, and reviewing the names in an interview transcript is a condition before you build a script on it. This is the same rule we set out in Using AI for research without inheriting its mistakes.
Sensitive material stays outside all of this
Before you upload any clip to a tool that works over the internet, ask: is there anything in this file that should not leave your computer?
A testimony from someone you agreed not to identify, something said in a private context, material for a client under a confidentiality agreement. These do not go to an outside service. And if the whole archive is mixed together, a clearly named folder kept out of any automated indexing is the simplest answer.
When this is worth your time
Fully indexing three years of archive is a large project, and it is not where you start. Begin with three steps in this order:
1. From today onward. Every new shoot gets its card the same day. This is a five-minute habit that stops the problem recurring.
2. The last six months. Material you still remember, where adding the names and consent status is quick and easy.
3. The old archive on demand. Do not index all of it. Every time you look for something and find it, write its card. After ten projects most of what you actually use is indexed, and whatever you have not touched in three years does not need indexing.
What you gain after two months
The obvious gain is that you find your shots. There are three less obvious and more important ones:
- Your cutaways become available. Instead of filming fresh coverage for every video, you have a searchable stock, which multiplies the value of what we covered in Cutaways in Editing: When to Use Them and Why.
- Your shot list gets shorter. When you build a shot list from the script, you delete whatever you already own, and the shoot day gets lighter.
- Ideas come out of the archive. Reading descriptions of what you filmed over two years reveals lines you had not noticed, and subjects whose material you already hold.
The rest of it is drive management rather than indexing: how many copies you keep and where, and what we explained in Batteries and memory cards: how to avoid losing a day's filming applies to the archive as much as to the card.
The whole thing in one line
The tool writes what is in the picture and in the sound; you write the date, the names, the consent and the judgement. Whoever splits the work that way finds a shot in a minute, and whoever waits for a tool that does everything keeps buying clips he already owns better versions of. This is the same logic that governs every smart editing tool, as in AI-assisted editing: what actually saves you time?.