AI Autofill: tag every shot with your own fields, automatically

StoryFolder — 2026-09-02

Part of: StoryFolder 0.4: AI Autofill, local transcription, and an assistant that can see your library

Last updated: 2 September 2026.

AI Autofill looks at one frame from each shot and writes a value into the shot metadata fields you switched it on for. New in StoryFolder 0.4. Every field ships with the switch off, so nothing is filled until you turn something on, field by field. Dropdown, Rating, Checkbox and Number answers are held to the domain you defined. Tag and Text are open on purpose and can come back with something you never listed. It is a StoryFolder Pro feature, it sends frames off your machine to a hosted vision model, and it writes into the cell with no restore point behind it, so reviewing what it wrote is your job and not the app's.

What can AI Autofill fill?

Shot fields you already created and explicitly enabled, of six types: Text, Tag, Dropdown, Checkbox, Rating and Number. It never creates a field, it never touches a video-level field, and it never fills a field whose Autofill switch you have not turned on. Dropdown, Rating, Checkbox and Number answers must fit the domain you defined; Tag and Text accept anything.

That split is the fact to carry away. A Dropdown cannot come back with an option you did not write: the answer is checked against your option list before it is accepted, and a near-miss is corrected to your spelling rather than added as a new option. A Number takes any finite number and refuses anything that is not one, so a tick or a list of tags cannot quietly become a quantity. A Tag field can grow, because a tag vocabulary that could only reuse what already exists would never let the model add the one you have not thought of yet. And a Text field has no vocabulary to fail against at all, which is why a Job# in a Text field with Autofill enabled is where a plausible wrong answer has nothing to stop it.

Video-level fields are a separate schema and are out of scope by design. There is no Autofill column in Settings → Video Metadata, because a client name and a job number are not in the picture. Video metadata and the rest of the library controls are their own thing.

Filling values and inventing fields are also different features. AI Discover fields takes a description of the work you do and proposes field names, types and options. It writes no values, and anything you accept from it arrives with its Autofill switch off.

How do your field rules shape the result?

The field's name, its type, its option list and the description you typed into it are what the model is asked for, and the domain table is what StoryFolder will accept back.

Take a four-field schema: Shot Size as a Dropdown with your own options, Location as a Tag, Description as Text, Usable as a Checkbox. For each shot, the run sends that shot's frame plus the video's title, description and any video-level fields you filled in, plus whatever you typed into the run's own instructions box. Each field goes along with its type and its options, and the request tells the model to leave a field out rather than guess when the frame does not support an answer. What comes back is checked against that field's domain before it becomes a value, and a one-line field description is the cheapest steering available: Shot Size with the description "framing of the main subject" produces different answers from Shot Size with nothing.

Passing that check is not the same as being right. A Dropdown answer that is one of your options is accepted whether or not it is the correct one of your options, and this is the failure mode you will actually meet. What "good" means for AI video tagging works through a real case of an accepted, valid, visually wrong value and what to do about it. How AI uses categories you already defined covers designing the option list so there are fewer of them.

How do you set it up without overwriting metadata you already have?

Enable one or two fields, leave Only fill empty fields on, run it on a small selection, and look at what came out before you widen the scope.

1. Open Settings → Shot Metadata. Custom fields are Pro, so this is where the plan matters.
2. Create or pick a field, turn on its Autofill switch in the Autofill column, and write it a one-line description.
3. Leave Autofill on import off. It is off by default, and turning it on before you trust the schema means a forty-file drop becomes hundreds of model calls with nobody watching.
4. In the board or the library inspector, select a handful of shots and choose Autofill with AI. The action stays hidden until at least one field has its switch on, so if you cannot find it, step 2 has not happened yet.
5. In the modal, check the field list, pick the scope (this shot, the shots you selected, or all of them), leave the mode on Only fill empty fields, and read the count line before you commit. It names fields and shots, and the button repeats the shot count.
6. Look at the results, fix what is wrong by hand, and only then widen.

The run itself is a queued job owned by the worker, so it keeps going if you navigate away, and it survives a restart. Progress shows per field on the card and in the inspector rather than as one bar for the whole board.

What can go wrong even when a value passes validation?

Four different things, and they do not look alike from the outside. A value can be correct. It can be inside your domain and visually wrong. It can be rejected for falling outside the domain, in which case nothing is written and nothing tells you so. Or the call can fail, which is retried for a network or server error and not retried when the model returns a refusal.

The middle two are the ones to plan around. A rejected value and a model that declined to answer both look like an empty cell, so an empty cell is not evidence about which happened. And an accepted-but-wrong value is indistinguishable in the cell from one you typed: there is no marker in the UI saying a value came from Autofill.

Reviewing a large board is a filtering job. On a 461-shot board, use each closed field's own options as a review queue: filter to Shot Size = Wide, look at that set, move to the next option. Open Text and Tag fields have no queue to filter into, so take a fixed sample across the beginning, middle and end plus the shots most likely to defeat a single frame, which are the dark ones, the transitions and the graphics. Correct what you find in place.

There is no restore point to fall back on. Autofill writes into the video directly and never mints one, so version history does not cover a run, and the only real guard is that Only fill empty fields never touches a cell that already has something in it.

Where does the frame go?

Out. StoryFolder encodes one frame per shot and posts it to StoryFolder's own server, which passes it to a hosted vision model along with your field names, your field descriptions, the run's instructions and the video's title, description and video-level fields. The video file itself is never uploaded.

So an Autofill run needs an internet connection every time, and the local story belongs to the feature next door. Transcription runs on your machine with no upload at all; Autofill is the opposite in every respect that matters here, including the plan it sits on. The privacy help page is the current authority.

Autofill is Pro, and the server enforces it rather than the client asking nicely. A non-Pro account meets an upgrade prompt in the run modal instead of a failed job.

When should you leave Autofill switched off?

For anything the frame does not contain, anything with contractual weight, and any field whose options you have not yet argued with yourself about.

A model can look at a shot and tell you it is B-roll. It cannot look at a shot and tell you who signed a release, which job number this is, or which of two clients booked the day. Those facts live in your head or in an email, and the honest route for them is a video-level field typed once, or copy and paste across shots where a run of shots shares a value.

Autofill is also not search. It writes values that make search work later; it does not go and find things. And on a schema you have not tested, running it wide is how you get 461 shots of confidently wrong metadata with no undo. Enable little, run small, widen when it has earned it. The end-to-end logging workflow is the longer version of that argument.

FAQ

Does AI Autofill work with my own metadata fields? Yes, that is the whole feature. It fills the shot fields you created and enabled, of six types: Text, Tag, Dropdown, Checkbox, Rating and Number. Custom fields are Pro.

Will Autofill overwrite metadata I already entered? Not in the default mode. Only fill empty fields is on by default and never touches a cell with something in it; choosing the overwrite mode shows a counted warning first.

Can I undo an Autofill run? No. Autofill creates no restore point, which is why the advice is to run it on a small selection and review before widening.

Does Autofill upload my video? No. It sends one frame per shot, plus your field names and instructions and the video's title, description and video-level fields, to StoryFolder's server and on to a hosted vision model.

Can Autofill fill Client or Job at video level? No. It works on shot fields only, and Settings → Video Metadata carries no Autofill control at all.

Why can I not see the Autofill action? It stays hidden until at least one shot field has its Autofill switch on. Turn one on in Settings → Shot Metadata.

How do I review a 461-shot board? Filter each closed field by each of its options in turn and inspect those queues, then take a fixed sample of the open Text and Tag results plus the dark, graphic and transition shots. Do not scroll the whole board.

Is AI Autofill free? No, it is on StoryFolder Pro, and the server enforces it. Transcription, by contrast, is free.

← All articles