Feature · All agencies

How PhotoTager prompts the AI to write stock photo metadata

A stock buyer types a few plain words and scans a page of thumbnails. The prompts behind PhotoTager are written for that moment: say what is shown, put the words buyers type first, and never guess. This page explains each rule, including what the prompts do not do.

What one prompt asks for

01 A title, an Adobe Stock title and a description Each has its own length range, set in your template. The Adobe Stock title is capped at 70 characters in the prompt and is written to be more descriptive than the plain title.
02 A ranked keyword list The list runs from the most obvious keywords to the most speculative. Some prompts also ask the model to score each keyword from 0 to 1 for relevance and use the scores to sort.
03 Suggested keywords for Adobe Stock Some prompts also ask for a second list of broader synonyms and related ideas, kept apart from the main keywords.
04 Your tone, style and word lists Your template sets the tone and style, whether titles use Title Case, keywords that must be used exactly as typed, and words that must never appear.

Who, what, where and when

  1. 1
    Who and what come from the photoThe model is told to describe the main subject, what people are doing, the size of a group, age range and the objects in the scene. It has no way to know a person’s name.
  2. 2
    Where and when come from youA camera cannot show that a lake is Lake Bled or that a market is Diwali in Jaipur. The project’s context note goes into the prompt as the “topic”, and the model is told to work it into the title and description. In the keyword order below, the topic comes first.
  3. 3
    No note, no guessingIf the note is empty or unclear, the prompt says to base the answer on the image alone, and not to assume anything beyond what the note says. A note in another language is translated to English first.
  4. 4
    Each run looks at four anglesThe prompt picks four of ten focus areas at random, such as setting, lighting, composition, clothing or relationship, so titles in a batch do not all read the same.

Buyer search terms first

  1. 1
    Write what a buyer would typePlain nouns and verbs, simple language, the specific subject and the action. The prompts say to write for buyers of stock photography and to keep the vocabulary simple.
  2. 2
    Describe, do not advertiseTitles and descriptions should state the content directly. The prompts ban filler such as “heartwarming”, “showcasing”, “vibrant”, “ideal for” and “this image captures”, and ask for no dashes, colons or similar punctuation in titles and descriptions. You can add your own banned words in the template.
  3. 3
    No repeated or near-duplicate keywordsThe prompts ask for no duplicates, and in some versions no variants of one word, for example both “backpack” and “backpacks”.
  4. 4
    Only what is visible or givenKeywords should describe what is clearly visible or what your note says. Low-confidence guesses go to the end of the list.

Titles and descriptions by agency

WhatWhat the prompt doesAgency limit it must fit
Adobe StockWrites a separate Adobe Stock title, up to 70 characters, more descriptive than the plain title.Titles up to 200 characters, about 70 recommended. Up to 49 keywords.
ShutterstockWrites a full-sentence description in plain text. Keep it short enough for the field.Descriptions up to 150 characters, at least 5 words in the title, 7 to 50 keywords.
iStock and Getty ImagesWrites the same title, description and ranked keywords. A separate step matches them to the iStock vocabulary.Up to 50 keywords, chosen from iStock’s own vocabulary.
AlamyWrites a factual description that can serve as the caption, and keywords that become the tags.Captions up to 150 characters, 5 to 50 tags.
Pond5Uses the same title, description and keywords.Check Pond5’s own rules before upload. PhotoTager does not enforce a Pond5 limit.

Keyword order: most important first

01 The order the prompt asks for In the main version: words from your note, then words from the generated title and description, then basic, easily recognised objects, then specific subjects and actions.
02 Technical terms stay few At most two terms such as “close up”, “wide shot”, “aerial view” or “natural light”. For video, terms such as “slow motion” and “time lapse” play the same role.
03 People terms come last If people are in the photo: group size, a broad ethnicity term only where it is clearly visible, an age group and a simple action word. If nobody is in it, the prompt says to leave people terms out.
04 Why the order matters on Adobe Stock Adobe tells contributors that the first 10 keywords carry the most weight. Other agencies do not publish an exact ranking rule, so we do not claim one.

What the AI should never write, and what it cannot catch

01 Rules in the prompt No filler or ad copy, no assumptions beyond your note, no punctuation agencies dislike in titles, and none of your own excluded words. The editorial prompts add: do not guess names, places or dates.
02 Brand names and trademarks The built-in prompts do not carry a brand list, so a photo of a phone can come back with the brand name. Add brands, logos and characters to your template’s excluded words, and check the metadata with the free checker, which flags known brand and character names, before you submit.
03 People’s names and releases The AI uses a name only if you put it in the note, and it cannot know whether you hold a model or property release. That is your decision. Name people only in editorial files, and check the release status of every recognisable person and property.
04 Claims nobody checked The prompts ask for what is visible and what your note says. They do not verify your note, so a wrong place name in the note becomes a wrong place name in the keywords.

Editorial format

  1. 1
    Turn on Editorial in the templateThe built-in editorial prompt asks for the Adobe Stock title and the description in the form “Location - Date - Description”, with the city, state or province and country, a full date, and a short factual description.
  2. 2
    Give it the place and the datePut the place, event and date in the project note. The prompt does not tell the model to invent them.
  3. 3
    Factual and neutralJournalistic language, no opinions, and factual keywords.
  4. 4
    Your own prompt in editorial modeIn the Playground, the Editorial switch adds a rule to your prompt: never guess names, places or dates, and when your note gives both the place and the date, start the description with “City, Country - MONTH DAY, YEAR:”, for example “London, UK - APRIL 27, 2019:”.

Where your context note and template fit

  1. 1
    The template sets the rulesLengths, tone, style, Title Case, keywords to use and words to avoid are filled into the prompt from the template you pick for the project.
  2. 2
    The note sets the factsThe project’s context note is inserted as the topic. It applies to every file in the project, so keep one shoot per project.
  3. 3
    Both go to the model with the imageThe model sees the picture, the template rules and the note together, and returns a structured answer that PhotoTager stores as the title, description and keywords.
  4. 4
    A custom prompt replaces the built-in oneIf your template has a custom prompt, it is used as you wrote it, followed by your context note. The template’s tone, length and word-list settings are not added to it, so put what you need in the prompt itself.

Write, test and save your own prompt

  1. 1
    Start from a starter promptThe Prompt Playground has starter prompts for stock SEO, editorial and short minimal metadata. Edit one or write your own, up to 6,000 characters.
  2. 2
    Run it on one photoPick a model from OpenAI, Anthropic Claude or Google Gemini, upload one photo, add a context line of up to 500 characters if you want, and run it. Change one thing at a time, then run again.
  3. 3
    Compare modelsThe same prompt gives different wording on different models. Run two or three on the same photo before you decide.
  4. 4
    Say how much you wantThe Playground always asks for a title, a description and a keyword list, most relevant first. Your prompt decides the lengths, the number of keywords and the style.
  5. 5
    Save it as a templateSign in and save the prompt. It becomes a template with the same model and the same wording, so a batch runs what you tested. Models marked Thinking take longer and cost a multiple of a standard run per image.

Check the output before you submit

01 Read the first 10 keywords On Adobe Stock these matter most. Move any weak word out of that group.
02 Check names, places and dates Check them against your note and your memory of the shoot.
03 Run the checker The free checker compares titles, descriptions and keywords with each agency’s limits and flags known brand names. Do this once before you export.

Questions

Can I trust AI keywords for stock photos?

Trust them as a first draft that you review. The prompts ask for visible content and for your note, but a model can still misread a photo, and it cannot know a place you did not tell it.

Do the prompts avoid brand names and trademarks?

Not by themselves. The built-in prompts do not contain a brand list. Add brands to your template’s excluded words, and use the free checker before you submit.

Why is the Adobe Stock title different from the description?

Adobe Stock has its own title field with a recommended length of about 70 characters. The prompt writes a plain title and a more descriptive Adobe Stock title.

Does PhotoTager use the same prompt for photos and video?

The structure is the same: title, description, ranked keywords. The video prompts also look at motion and camera work and use video terms such as slow motion and time lapse.

Can I see the full prompt?

No. The built-in prompts are part of the product, so this page describes what they do. Your own prompts in the Playground are fully visible to you.

Will my custom prompt use my context note?

Yes. The note is added after your prompt as “Additional context from the photographer”. In editorial mode the editorial rules are added too.

Is testing a prompt free?

The Playground gives you free runs without an account, and the page shows how many are left and what a run costs after that.