[AEM Assets]{class="badge positive"}

Brand-Aware Metadata best practices: a WKND tutorial brand-aware-metadata-best-practices

Complete, accurate metadata is what makes assets findable, reusable, and accessible. Producing it takes real effort. For every asset, someone has to understand the content, know the business context, and fill in descriptions, classifications, and other fields in a way that follows your brand’s rules and taxonomy. Generic AI capabilities describe what is visible in an asset, but they cannot know your brand voice, your campaigns, your approved vocabulary, or details that exist only in other systems.

Brand-Aware Metadata (BAM) in Adobe Experience Manager Assets generates values for your own metadata properties, such as free-text descriptions, alt text, tag-based classifications, and dropdown selections, using brand context, approved values, and existing metadata. The quality of the generated metadata, however, depends directly on how you set it up.

This article applies Adobe’s recommended best practices to a complete, end-to-end example for WKND, a fictitious adventure-travel brand. You prepare the metadata model, write and refine prompts, add brand and business context, and validate the results before you scale. By the end, four brand-specific properties are generated automatically for every asset in a campaign folder.

As part of this tutorial, you configure and generate the following properties for all assets in a campaign folder:

Property
Field type
Why it is hard to fill in
Alt Text
Free text
Needs the brand’s voice, not a generic caption.
Activity Type
Controlled vocabulary (tags)
Needs WKND’s own terms, including niche ones.
Region
Controlled vocabulary (dropdown)
Cannot be read from the pixels at all.
Campaign
Controlled vocabulary (dropdown)
Pure business context.

Along the way, you use five BAM capabilities: Prompt Playground, Prompt Enhancer, brand prompts, controlled vocabulary, and referenced properties.

NOTE
This article builds on Generate Brand-Aware Metadata, which describes each BAM capability in detail.

Best practices at a glance best-practices-at-a-glance

The following table lists Adobe’s recommended best practices for Brand-Aware Metadata, in the order in which you first apply them in this tutorial.

Best practice
Where you apply it
Scope each controlled-vocabulary field to only its relevant set of allowed values.
Step 2: Create the tag hierarchy and Step 3: Add the fields to a metadata schema
Use brand prompts for global context, and scope them to processing profiles for folder-level control.
Step 4: Apply the schema and a processing profile, Step 7: Add a brand prompt, and Step 10: Derive the campaign
Structure prompts one-to-one with properties, and organize them around the property, not individual values.
Step 5: Create a property prompt
Iterate in Prompt Playground before you touch real assets, then use Prompt Enhancer to tighten the prompt.
Step 5: Create a property prompt and Step 6: Refine the prompt
Constrain output with controlled vocabulary, and start with as little definition as possible.
Step 8: Constrain with controlled vocabulary and Step 9: Add referenced properties
Use referenced properties to build on existing metadata.
Step 9: Add referenced properties
Build a quality gate before scaling: sample, review in the quality report, and route by confidence.
Step 11: Validate on a sample, then scale
Use BAM for deterministic fields too, but weigh the credit cost.
Additional best practices

Example scenario: WKND example-scenario

The WKND content team manages thousands of trip photos. Photographers upload assets from Adobe Bridge, and trip details arrive from Adobe Workfront. Before the marketing team can find and use an asset, it needs alt text, an activity type, a region, and a campaign. Today, a librarian fills in those fields manually.

This tutorial uses two sample assets, both stored in /content/dam/wknd/campaigns/winter-backcountry-2027:

Asset A
Asset B
File
WKND_ALP-BC-07_0412.jpg
WKND_TET-SB-02_0198.jpg
Image content
Three skiers on climbing skins, crossing a snowy ridgeline at sunrise, wearing avalanche packs.
A rider carving down an untouched powder slope, with no lifts or other gear visible.
Shoot brief (from Workfront)
“Guided three-day ski-touring itinerary, Mont Blanc massif, intermediate pace.”
“Two-day splitboard weekend, Teton Range, Wyoming. Intermediate.”
Why it is useful
Clear visual signals.
Visually ambiguous, so it needs context.
NOTE
The generated values shown in this article are representative examples. Your results can vary because AI output is not deterministic.

Prerequisites prerequisites

Ensure the following:

  • Access to Adobe Experience Manager Assets as a Cloud Service with Brand-Aware Metadata enabled.
  • Administrator permissions, or the permissions required to create tags, edit metadata schemas, create processing profiles, and access Prompt Editor.
  • The folder /content/dam/wknd/campaigns/winter-backcountry-2027 with a few sample assets.

Step 1: Review the baseline metadata for an asset review-baseline-metadata

Before you configure BAM, look at what you start with, so that you can measure the improvement later.

  1. In the Assets user interface, navigate to the winter-backcountry-2027 folder.
  2. Select Asset A and click Properties.
  3. Review the generic Smart Tags already applied to the asset. Typically, they look like snow, mountain, skiing, person, and sky.
  4. Note that the alt text, activity type, region, and campaign fields are not yet available in the asset properties. You add these fields in the next steps.

To fill in those fields manually, a librarian reads the shoot brief, opens the tag picker, finds the right activity, picks a region from the list, writes alt text, and picks a campaign. That work is repeated for every asset.

Checkpoint, the baseline: Generic tags describe pixels. None of the four WKND fields are populated, and filling them in manually takes several minutes per asset.

Step 2: Create the tag hierarchy for Activity field create-tag-hierarchy

Execute the following steps to create a tag hierarchy for the Activity field.

  1. Navigate to Tools > General > Tagging.

  2. Click Create > Create Namespace, enter the following values, and click Create:

    • Title: WKND
    • Name: wknd
  3. Select the WKND namespace, click Create > Create Tag, and create a tag with the Title Activity and the Name activity. This tag is the root of the activity branch.

  4. Select the Activity tag and create the following category tags under it. Then select each category tag and create its child tags:

    table 0-row-2 1-row-2 2-row-2 3-row-2 4-row-2 5-row-2
    Category tag Child tags
    Snow Sports Backcountry Skiing, Resort Skiing, Snowboarding, Splitboarding
    Climbing Alpine Climbing, Sport Climbing, Bouldering
    Water Sports Surfing, Sea Kayaking, Stand-Up Paddleboarding
    Cycling Gravel Riding, Mountain Biking
    Camp & Trek Backpacking
  5. Publish the tags if your assets are also used on publish.

The result is a branch at /content/cq:tags/wknd/activity. For detailed instructions on the tagging console, see Administering tags.

TIP
Give each tag a clear, descriptive title. Modern models map content to a well-named set of values accurately, often without any additional definition.

Step 3: Add the brand fields to the metadata schema add-fields-to-metadata-schema

In this step, you add the four WKND fields to the metadata schema, plus a field for the shoot brief that Workfront provides.

IMPORTANT
Configure the schema fields and their allowed values before you create prompts. A property prompt loads its properties and controlled vocabulary values using the metadata form.

Create the WKND schema create-wknd-schema

  1. Navigate to Tools > Assets > Metadata Schemas.
  2. Select the default schema form, click Copy, and name the copy WKND. Alternatively, click Create to start a new form. For details, see Add a metadata schema form.
  3. Select the wknd form and click Edit.
  4. In the Metadata Schema Form Editor, click + to add a tab and name it WKND. Adding a dedicated tab keeps the brand fields together on the asset Properties page.

Next, drag the following components from the Build Form tab to the WKND tab:

NOTE
Use a custom namespace prefix such as wknd: in Map to property. A custom namespace keeps brand-specific properties separate from standard properties and makes them easy to identify when you configure prompts and referenced properties.

Add the Alt text field add-alt-text-field

  1. Drag a Single Line Text component to the tab.

  2. Click the field and configure the following options using the Settings tab:

    • Field Label: Alt text
    • Map to property: ./jcr:content/metadata/Iptc4xmpCore:AltTextAccessibility
    • Description: Accessible description of the image, under 125 characters.

Add the Activity type field add-activity-type-field

  1. Drag a Standard Tags component to the tab.

  2. Click the field and configure the following options using the Settings tab:

    • Field Label: Activity type
    • Map to property: ./jcr:content/metadata/wknd:activity
    • Root Path: /content/cq:tags/wknd/activity

Setting the root path is the key best practice for tag fields. The field, and BAM, then see only the activity branch rather than the entire tag tree.

Add the Region field add-region-field

  1. Drag a Dropdown component to the tab.

  2. Click the field and configure the following options using the Settings tab:

    • Field Label: Region
    • Map to property: ./jcr:content/metadata/wknd:region
  3. Under Choices, Select Add Manually and click Add Choice for each option and enter the following option text and values:

    table 0-row-2 1-row-2 2-row-2 3-row-2 4-row-2 5-row-2 6-row-2 7-row-2 8-row-2
    Option text Value
    European Alps european-alps
    Rocky Mountains rocky-mountains
    Pacific Northwest pacific-northwest
    Patagonia patagonia
    Scandinavia scandinavia
    Iceland iceland
    New Zealand new-zealand
    Southeast Asia southeast-asia

Add the Campaign field add-campaign-field

  1. Drag a Dropdown component to the tab.

  2. Click the field and configure the following options using the Settings tab:

    • Field Label: Campaign
    • Map to property: ./jcr:content/metadata/wknd:campaign
  3. Under Choices, Select Add Manually and click Add Choice for each option and enter the following option text and values:

    table 0-row-2 1-row-2 2-row-2 3-row-2 4-row-2
    Option text Value
    Winter Backcountry 2027 winter-backcountry-2027
    Summer Surf Camps 2027 summer-surf-camps-2027
    Gravel & Grit 2027 gravel-and-grit-2027
    Everyday Explorer (evergreen) everyday-explorer

Add the Shoot brief field add-shoot-brief-field

  1. Drag a Single Line Text component to the tab.

  2. Configure the following settings:

    • Field Label: Shoot brief
    • Map to property: ./jcr:content/metadata/wknd:shootBrief
  3. In Rules tab, Select Disable Edit, because Workfront writes this value upstream.

Adding this field to the form is optional. BAM can reference a property even when it does not appear on the form. Displaying it, however, helps reviewers see the context that BAM used. Make a note of the exact property name, wknd:shootBrief, because you type it when you configure referenced properties in Step 9.

Click Save to save the schema.

Why this matters: A field scoped to a small, relevant set of values is faster and less error-prone for anyone picking a value manually. It also makes BAM more accurate, because the model chooses from a short list instead of reasoning over an entire enterprise taxonomy.

Step 4: Apply the metadata schema and a processing profile to the campaign folder apply-schema-and-profile

When you apply metadata schema to a folder, the fields included in the metadata schema display on the asset Properties page for all assets within that folder. Once you have created the WKND metadata schema and configured the fields, you must apply the metadata schema to the /content/dam/wknd/campaigns/winter-backcountry-2027 folder.

Apply the metadata schema apply-metadata-schema

  1. Navigate to Tools > Assets > Metadata Schemas.
  2. Select the wknd schema and click Apply to Folders.
  3. Select /content/dam/wknd/campaigns/winter-backcountry-2027 and click Apply. If the folder already uses another schema, click Overwrite.
  4. Open Asset A, click Properties, and confirm that the WKND tab displays the new fields.

For details, see Apply folder-specific metadata.

The processing profile determines which BAM prompts run for the assets in a folder. You can associate multiple property prompts to a processing profile and similarly associate a processing profile with a brand prompt as well. When the brand prompt and the property prompt have the same processing profile assigned to them, the metadata for a field is automatically generated based on the property prompt as well as the brand prompt.

Create and apply a processing profile create-processing-profile

  1. Navigate to Tools > Assets > Processing Profiles.
  2. Click Create, name the profile WKND Winter Backcountry 2027, and click Save. For details, see Create a standard profile.
  3. Select the profile, click Apply Profile to Folders, select the winter-backcountry-2027 folder, and click Apply.

Only one processing profile can be applied to a folder, and it applies to all subfolders unless a subfolder has its own profile. For details, see Use processing profiles to process assets.

TIP
Plan your folder structure around metadata context. Because each processing profile can carry one brand prompt, the folder hierarchy becomes the mechanism that routes context to assets. For example, a /campaigns/winter-backcountry-2027 folder and a /campaigns/summer-surf-camps-2027 folder can each carry their own campaign context.

Step 5: Create a property prompt and test it in Prompt Playground create-property-prompt

Start with one property and a short, plain-language prompt. Do not try to populate several fields from one prompt.

  1. In Assets, switch to Admin view and navigate to Assets > Brand-Aware Metadata.

  2. Click Create property prompt.

  3. Configure the following fields:

  4. In the Prompt field, enter:

    code language-none
    Write alt text for this image. One sentence, under 125 characters.
    Describe what the people are doing and where.
    Don't start with "Image of" or "Photo of".
    
  5. In the Prompt Playground section, select Asset A as a preview asset, then click Generate preview output.

  6. Review the generated value and its confidence score.

  7. Click Save.

NOTE
  • The Metadata form list displays only metadata forms defined in Admin view.
  • The Description field helps you organize prompts but does not affect generation. Put all instructions in the Prompt field.
  • Prompt Playground previews results without writing anything to your assets and without consuming generation credits.

Example output: “A breathtaking view of skiers trekking up a stunning snow-covered ridge at sunrise.”

Checkpoint, accurate but generic: The caption is correct and took a single prompt. However, it sounds like every other AI caption. It uses hype words and says nothing about WKND’s voice.

Step 6: Refine the prompt with Prompt Enhancer refine-prompt-with-enhancer

A first draft is rarely your best prompt. Prompt Enhancer rewrites a plain-language prompt into a tighter, better-structured version for you to review.

  1. Navigate to Assets > Brand-Aware Metadata > Prompt List. Select the WKND alt text prompt and click Edit.

  2. Click Suggest enhancements.

  3. Review the suggested rewrite and its confidence metrics. You can apply it, edit it, or discard it. Nothing changes until you apply it.

    An example of a suggested rewrite:

    code language-none
    Write alt text for this image, for screen-reader users.
    - Format: one sentence, under 125 characters.
    - Content: say what the people are doing and where.
    - Style: start with the subject. Don't begin with "Image of" or "Photo of".
    
  4. Click Apply enhanced prompt.

  5. Click Generate preview output again. In the Prompt Playground section, select preview assets to evaluate how the prompt performs against sample assets.

  6. Click Save.

Example output: “Skiers trek up a snow-covered ridge at sunrise.”

Checkpoint, tighter but still not WKND: The caption is cleaner and follows the format rules more reliably. Prompt Enhancer improves how the prompt is written, but it does not know your brand. A brand prompt provides that context.

TIP
  • Run Prompt Enhancer on each property prompt as you write it, then confirm the results in Prompt Playground before you scale.
  • Keep prompts concise. Long prompts consume more input tokens and are more likely to contain contradictory instructions that reduce accuracy.

Step 7: Add global context with a brand prompt add-brand-prompt

A brand prompt holds context that applies to all property prompts in a processing profile, so you write it once. Use a brand prompt for any higher-level context, such as voice, global rules, naming conventions, or campaign details.

  1. Navigate to Assets > Brand-Aware Metadata > Brand Prompts and click Create brand prompt.

  2. In Prompt name, specify WKND Winter Backcountry 2027.

  3. In the Prompt field, enter:

    code language-none
    You are generating metadata for WKND, an adventure-travel brand that runs
    guided trips for people who want real outdoor experiences without the
    elitism. Voice: warm, plainspoken, inclusive. Never hyped.
    Avoid clichés and superlatives such as "breathtaking", "stunning", "epic",
    "majestic". Describe what people are doing, not how they look.
    Mention the key gear when it is visible.
    Always write in US English. Never invent place names or gear brands that
    aren't visible or provided.
    
    Folder context: assets in this folder support the Winter Backcountry 2027
    campaign, which promotes guided backcountry ski and splitboard trips for
    intermediate riders.
    
  4. Click Save.

  5. Associate the brand prompt with the WKND Winter Backcountry 2027 processing profile. A brand prompt has a one-to-one relationship with a processing profile, and processing profiles apply at the folder level, so this context applies to every asset in the winter-backcountry-2027 folder.

  6. Add the WKND alt text property prompt to the same processing profile. In the prompt list, select the prompt and click Add to Processing Profile. For details, see Apply AI metadata using processing profiles.

  7. Open the WKND alt text prompt, confirm that the brand prompt is selected, and click Generate preview output for Asset A. You have not changed the property prompt.

Example output: “Three skiers on climbing skins trek along a snowy ridgeline at sunrise, wearing avalanche packs.”

Checkpoint, same prompt, now on brand: The hype is gone. The caption is plain and specific, and it mentions the gear that matters. The property prompt did not change. The brand prompt did the work.

TIP
To see exactly what BAM sends to the model, click the more options (…) icon in the prompt playground for any of the preview assets and select View prompt. The brand prompt is prepended to the property prompt, and the file name and file path are included by default.

Step 8: Constrain a field with controlled vocabulary constrain-with-controlled-vocabulary

For a governed field, such as a tag or dropdown field, bind the prompt to a controlled vocabulary, so that BAM selects only from your approved values rather than generating free text. Free text does not match your taxonomy and breaks faceted search and reporting. BAM matches each generated value to the value stored in AEM for you.

Start with as little definition as possible. Then run the prompt to learn how much help the model actually needs.

  1. Click Create prompt and configure the following fields:

    • Prompt name: WKND Activity Type
    • Metadata form: WKND
    • Metadata property: wknd:activity

    Because Activity type is a tag field, the editor displays the controlled vocabulary builder with a tag selector instead of a free-text output.

  2. Select whether the property returns a single value or multiple values. For this example, select a single value.

  3. Add the approved values. The tag selector shows only the tags under the root path that you configured in Step 3. Select the leaf tags under Snow Sports, Climbing, Water Sports, Cycling, and Camp & Trek. Select the option to return Bottom-level tags only, so that BAM does not return a category tag such as Snow Sports. You can also click Add all to add all tags created under Activity.

  4. Leave the value definitions empty for now. In the Additional text to append field, specify:

    code language-none
    Select the activity shown in this image.
    

    The Controlled vocabulary field displays the values based on what you specify in the Additional text to append field and what tag values you select in the Value Definitions section.

  5. In Prompt Playground, select Asset A and Asset B, then click Generate preview output.

    A highlighted result in the preview indicates a value that matches the controlled vocabulary. A value that is not highlighted is free text and not the values that match the defined controlled vocabulary.

  6. Click Save.

Example output: Asset A returns Backcountry Skiing, a valid WKND value with no additional guidance.

Asset B returns Snowboarding. It is close but not what WKND wants. The image shows a rider carving down a powder slope, and a split snowboard looks like any other snowboard from this angle. The pixels do not contain the answer.

Checkpoint, pixels alone hit a ceiling: For Asset B, the model chose a valid and reasonable value, but not the value that WKND wants. A split snowboard looks like any other snowboard from this angle. Describing the image in more detail does not fix this, because the missing information is not in the image.

IMPORTANT
If you add new tags or dropdown options later, add them to the prompt’s controlled vocabulary as well. Existing prompts do not pick up new values automatically.

Step 9: Add context with referenced properties add-referenced-properties

BAM runs after all other metadata extraction is complete. By the time BAM runs, the asset already carries embedded metadata, such as fields set in Adobe Bridge, and metadata written by upstream systems, such as Workfront. Every WKND shoot starts as a Workfront project, so each asset arrives with the wknd:shootBrief property. A referenced property lets a prompt read that existing metadata.

Refine the Activity type prompt refine-activity-type

  1. Open the WKND Activity Type prompt.

  2. In the Referenced properties field, type wknd:shootBrief. Enter the exact property name used in Map to property. You can reference a property even if it does not appear on the metadata form.

  3. Specify the following in the Additional text to append field:

    code language-none
    Select the activity shown in this image.
    Use the shoot brief as a supporting signal when the image alone is
    ambiguous. If the brief and the image disagree, trust the image.
    
  4. The model can still confuse two adjacent values, so add a short definition to only the value that needs it. In the controlled vocabulary, add the following value definition for Splitboarding. There is a separate text box adjacent to each controlled vocabulary option, the Value Definition that defines when to prefer any particular option over another in case of a confusion.

    code language-none
    Backcountry snowboarding on a board that splits into two skis for the
    climb. Choose this when the shoot brief mentions a splitboard.
    

    Write definitions that explain how similar values differ, rather than repeating general context.

  5. In Prompt Playground, select Asset B and click Generate preview output.

  6. Click Save.

Example output: Asset B returns Splitboarding.

TIP
You can also click Suggest enhancements to improve value definitions, not only the main prompt.

Create the Region prompt create-region-prompt

  1. Click Create prompt and configure the following fields:

    • Prompt name: WKND Region
    • Metadata form: WKND
    • Metadata property: wknd:region
    • Referenced properties: wknd:shootBrief
  2. Select a single value, and add all eight regions from the dropdown.

  3. In the prompt field, enter:

    code language-none
    Select the region where this image was taken.
    Use the location in the shoot brief. Choose only from the available
    regions. If the brief gives no usable location, leave the field empty.
    
  4. In Prompt Playground, select Asset A and Asset B, then click Generate preview output.

  5. Click Save, and add the prompt to the WKND Winter Backcountry 2027 processing profile.

Example output:

  • Asset A: European Alps, because the brief mentions the Mont Blanc massif.
  • Asset B: Rocky Mountains, because the brief mentions the Teton Range, Wyoming.

Checkpoint, business context, not pixels: No image contains the words “Mont Blanc” or “Teton Range,” and a generic tagger cannot produce a region at all. BAM read both the pixels and the shoot brief, then mapped a place name to WKND’s own region list.

TIP
Treat referenced values as a supporting signal, not as the single source of truth. This matters most when the referenced data was itself generated by AI or is outside your control. The instruction “If the brief and the image disagree, trust the image” prevents an incorrect upstream value from overriding what is visible.

Step 10: Derive the campaign from folder context derive-campaign

The campaign is not in the image or in the file name. It comes from the brand prompt that is tied to the folder through the processing profile.

  1. Click Create prompt and configure the following fields:

    • Prompt name: WKND Campaign
    • Metadata form: Wknd
    • Metadata property: wknd:campaign
  2. Select a single value, and add the four campaigns from the dropdown.

  3. In the Additional text to append field, specify:

    code language-none
    Select the campaign this asset supports. Use the campaign described in
    the brand context. If the asset doesn't fit it, select
    "Everyday Explorer (evergreen)".
    
  4. In Prompt Playground, select Asset A and Asset B, then click Generate preview output.

  5. Click Save, and add the prompt to the WKND Winter Backcountry 2027 processing profile.

Example output: Winter Backcountry 2027 for both assets.

Checkpoint, context that lives in the folder: If you put similar assets in a summer-surf-camps-2027 folder whose processing profile has its own brand prompt, the same campaign prompt returns a different response.

Step 11: Validate on a sample, then scale validate-and-scale

Generating metadata at scale is only as trustworthy as your review process. Do not go straight from a working prototype to your whole library. Build a quality gate first.

  1. Confirm that the brand prompt and the four property prompts (WKND alt text, WKND activity type, WKND region, and WKND campaign) are associated with the WKND Winter Backcountry 2027 processing profile.
  2. Start with a small, representative sample. Select about 20 existing assets in the folder, then click Reprocess Assets. For details, see Apply BAM to existing assets.
  3. Upload one new asset to the folder, and verify that it is processed automatically with the same profile.
  4. Open a processed asset, click Properties, and review the values on the WKND tab.
  5. In Brand-Aware Metadata Manager, open the quality report. Filter by date range and by lower confidence, so that you spot-check the assets the model was least sure about instead of reviewing every asset.
  6. Refine the prompts if required, and repeat the sample run.
  7. Get organizational sign-off. Treat approval of AI usage as a planned step, because it is frequently the real gate on going live. Start with the fields you are most confident in, and expand from there.
  8. Reprocess the rest of the folder.
NOTE
  • Quality report data is updated overnight and includes assets processed the previous day.
  • Processing overwrites existing values in the target properties. To keep existing values in a multi-value field, see Additional best practices.
  • Playground output sets your expectation for production. If a production run behaves differently from the playground, stop and reconcile the difference before you scale.

Compare the results compare-results

The following table compares Asset A before and after BAM:

Property
Manual or generic tagging
BAM with WKND context
Alt text
“Skiing on a snowy mountain,” or a caption full of hype words.
“Three skiers on climbing skins trek along a snowy ridgeline at sunrise, wearing avalanche packs.”
Activity type
“Skiing,” which is not a WKND value.
Backcountry Skiing
Region
Cannot be determined from pixels.
European Alps
Campaign
Cannot be determined from pixels.
Winter Backcountry 2027
Effort per asset
Several minutes by hand.
One configured prompt set, reused for every asset.

In summary:

  • One plain-language prompt per property replaces manual tagging for every asset. You wrote each prompt once, tested it in Prompt Playground, tightened it with Prompt Enhancer, and reused it across the folder.
  • The output is business context, not only a description of pixels. The brand prompt supplied WKND’s voice and campaign. The controlled vocabulary kept values within WKND’s taxonomy. The referenced property supplied the location and the niche activity that no image could show.

Additional best practices additional-best-practices

  • Organize prompts around the property, not individual values. Avoid value-centric rules, such as “if the product is a dress, tag it dress.” Write one prompt per property and express when each value applies through controlled vocabulary definitions. This approach keeps your prompt library manageable and helps the model distinguish adjacent values, such as a coat and a jacket.
  • Append values by referencing the property itself. BAM overwrites target properties by default. To add to a multi-value field, such as keywords or tags, reference the target property in the prompt and instruct the model to return the existing values plus the new ones.
  • Weigh the credit cost of deterministic fields. Some values are derivable from a rule, such as extracting a product code from a file name, mapping a folder path to a region, or applying a constant value. BAM can generate these values, but each generation consumes credits. Use BAM when the convenience of staying in the product outweighs the cost, and be mindful of large batches.
  • Iterate before you scale. Test in Prompt Playground, enhance, test again, then move to a small sample of real assets. Treat Prompt Playground as a mandatory first step, not an optional convenience.
recommendation-more-help
experience-manager-cloud-service-help-main-toc