ChatGPT keywords for stock footage: a prompt and a method

Which frames to give ChatGPT or Claude

Chat assistants read images. OpenAI's vision guide accepts PNG, JPEG, WebP and non-animated GIF, and its file uploads FAQ lists text files, spreadsheets, presentations and documents as supported types. Anthropic's guide for Claude accepts JPEG, PNG, GIF and WebP, and uses only the first frame of an animation. So the first step is yours: turn each clip into a few stills.

Any video player can save a frame. For a whole folder, FFmpeg is faster. This command saves one JPEG every four seconds, which gives five stills for a 20-second clip:

ffmpeg -i clip.mov -vf fps=1/4 -q:v 2 clip_%02d.jpg

Pick the frames with care, because they are all the model will know:

• Start, middle and end. Three frames show change: a subject entering, a drone revealing a coast, a sky going dark.
• Sharp, full-size frames. Anthropic says Claude can make mistakes on low-quality, rotated or very small images under 200 pixels. Skip blurred transition frames.
• Expect weaker results on fisheye shots. OpenAI says its model struggles with panoramic and fisheye images.
• One clip per message. Frames from two clips in one request produce one blended keyword list.

The model also cannot read your file. OpenAI says its model "doesn't process original file names or metadata", and Anthropic says Claude does not receive metadata from images. The place, the camera move and the frame rate have to be typed in, which is what the facts block of the prompt is for.

A ChatGPT prompt for Shutterstock, Adobe Stock and Pond5 keywords

This prompt works the same in ChatGPT and Claude. Fill the brackets, attach the frames in order, and send. The output fields are sized to fit all three agencies at once:

You are writing stock footage metadata for ONE video clip.
The attached images are still frames from that clip, in order:
the first image is the start of the clip, the last one is the end.

Facts I know about the clip (use them as given, do not change them):
- Location: [city, region, country, or "unknown"]
- Camera move: [static / pan left / tilt up / dolly in / drone flying forward]
- Speed: [real time / slow motion / time-lapse]
- Use: [commercial / editorial]
- Other: [species, event, anything the frames cannot show]

Rules:
1. Describe only what is visible in the frames or listed in my facts.
   If you are not sure of a place, a species or an object, leave it out.
2. No brand names, product names, logos, camera or drone models,
   people's names or other personal information.
3. English only. Use each word once, and only one form of a word
   (for example "run" or "running", never both).
4. Do not pad the list with loosely related words. Stop when the clip
   is described.

Return plain text in exactly this format:

DESCRIPTION: one factual sentence under 150 characters that says
who or what, doing what, and where. (For Shutterstock.)

TITLE: 40 to 70 characters, a short phrase describing the clip.
(For Adobe Stock and Pond5.)

KEYWORDS: 15 to 40 keywords, comma-separated, most important first.
The first 10 must cover the main subject, the action, the place
and the setting. Include the camera move and speed from my facts.

UNSURE: words you left out because the frames could not confirm them.

Two adjustments per agency. For Pond5, the first 20 keywords of the list match its 10 to 20 recommendation. For Adobe Stock, titles and keywords must match the language set in your contributor account; if it is not English, ask for a second version in that language. The background on each limit is in keyword limits by agency.

How to check the answer

OpenAI's own guide says its model "may generate incorrect descriptions or captions", and Anthropic says Claude "might hallucinate or make mistakes". Read each answer with the clip playing:

1. Read the UNSURE line first. It tells you where the model hesitated; settle each word there yourself.
2. Check places and species. A model that sees a stone bridge may name a famous one. Keep a place only if it is in your facts, and a Latin species name only if you can confirm it.
3. Strike brands and gear that slipped through, including phone or drone models in the description.
4. Look for stems and near-duplicates: sunset and sunsets, aerial and aerial view, city and cityscape side by side.
5. Count characters and keywords in the free title and keyword checker. Do not take the model's own count of the 150 or the 70 on trust.
6. Set the order for Adobe before you submit: it is locked while the clip is in moderation.

If the same mistake comes back clip after clip, fix the facts block or the rules instead of correcting each answer by hand. What a complete list looks like for common subjects is in stock footage keyword examples.

What a chat assistant cannot do for a batch

The method works. Its cost is the handling around it, and that cost grows with every clip.

At five frames per clip, OpenAI's cap of 80 files every three hours means 16 clips per three-hour window. The CSV route is covered in the Shutterstock CSV guide and the Adobe Stock CSV guide.

When a dedicated tool is worth it

The chat method makes sense when the volume is small or the clips are unusual. A dedicated tool starts to pay off when:

• you prepare more clips per week than you want to handle one by one in a chat window;
• you submit the same clips to several agencies and need a CSV or a filled form for each;
• you want frames extracted and limits enforced automatically, without counting characters;
• you want the file itself checked before a reviewer sees it.

What tools do with frames, and the errors they share with chat assistants, is explained in AI keywording for stock footage. Tools that handle video are compared on the same criteria, with their published prices, in keywording tools for stock footage.

Where RushPilot fits

RushPilot is a desktop app for stock video contributors, in private beta. It runs the loop on this page for a whole folder: it reads five frames of each clip, writes the title, description and English keywords to each agency's limits (in order of importance for Adobe Stock), checks common rejection reasons such as resolution, duration, visible logos and upscaling, and waits for your approval before it uploads to Shutterstock, Adobe Stock and Pond5.

It does not replace your knowledge of the shoot: a place or a species the frames cannot prove still needs your eye, and no tool can promise acceptance. It costs €0.05 to €0.09 per clip depending on the offer, with 25 clips free for new accounts; details on the pricing page. You can join the waiting list for the beta.