Skip to main content
AI Agents are reusable sets of instructions that tell Cernel’s AI exactly how to generate content for your products. Set up an agent once (define the prompt, style, and behavior) then assign it to categories in your taxonomy. When you run enrichment, agents generate attribute values for hundreds or thousands of products at once, consistently and on-brand.

How it works

An AI agent is a configured AI with a specific job: generate product descriptions, extract materials from images, classify products by gender, create meta titles, whatever content your products need. You build an agent by defining its prompt (the instructions the AI follows), selecting the attribute type it produces (Text, HTML, Single Select, Multi Select, External Data, or Media), and optionally referencing Data Sources for brand guidelines and context. Once created, you link the agent to an attribute on one or more categories in your Cernel Taxonomy. The agent then generates content for every product in those categories when you run enrichment. The key concept: agents are independent from categories. You create them once and reuse them wherever they’re needed. Change an agent’s prompt, and every category using that agent picks up the update.

Why you would use this

Write instructions once, generate everywhere

Instead of writing separate prompts for every category, create an agent and assign it wherever it applies. One “Product Description” agent can serve your entire catalog.

Keep your brand voice consistent

Agents reference your Data Sources (brand guidelines, tone of voice, product terminology). Every product gets content that sounds like your brand, not generic AI output.

Specialize by product type

Use different agents for different categories. A footwear description agent knows to mention cushioning and sole type. A clothing description agent focuses on fit and fabric. Same attribute, different expertise.

Iterate without re-doing work

Update an agent’s prompt and re-run enrichment to regenerate content with the new instructions. No need to reconfigure every category. The agent update cascades automatically.
Key concept: Agent. A configured AI with specific instructions, called prompts, that generates data for your products. Agents are created once and linked to attributes across categories in your Cernel Taxonomy.

Planning your agents

Before creating agents, think about what content your products need. A typical setup includes:
  • A product description agent (Text or HTML) for your main product copy
  • A meta title and meta description agent (Text) for SEO
  • A materials or features agent (Multi Select) for structured attributes
  • Specialized agents for specific product types (e.g., technical specs for electronics)
Each agent handles one attribute type and can be reused across many categories. Start with your most important content types and expand from there.

Creating an agent

1

Open the Agents page

In the sidebar, click Agents under the Tools section. You’ll see a table of your agents, each with its type, the categories and attributes it is linked to, and when it was created. Search by name, or sort by linked categories or creation date. Click an agent’s name to open it in the sandbox.
The Agents page showing a table with Agent, Linked Categories, and Created columns, a Search agents field, and agents such as Web Research (External Data), Product Description (HTML), and Product Image (Media/Image), each with a delete icon

The Agents page: every agent with its type, linked categories, and creation date

2

Click Create Agent

Click Create Agent to open the Agent Sandbox, the workspace where you build, test, and refine the agent in one place. The sandbox opens on a start screen that asks What should your agent do?
The 'What should your agent do?' start screen with a description box, a 'Select agent type (optional)' row of six type chips (Text, HTML, Single Select, Multi Select, External Data, Media/Image), and an 'Other options' section listing Use a template, Copy from existing, and Manual input

The start screen of a new agent: describe what it should do, optionally pin its type, or pick one of the other ways to begin

3

Choose how to start

Pick the way that suits what you already have:
  • Describe your intent (recommended): type a short description of what the agent should do, for example “Write a 150-word product description that’s professional but approachable, focusing on materials and fit”, and click Generate agent. The chat drafts a full agent with the right structure, references, and instructions.
  • Use a template: start from a ready-made agent in the template library. See Quick start with agent templates below.
  • Copy from existing: create a new agent that starts with the settings of one you already have.
  • Manual input: enter the name, type, and prompt yourself, with no AI assistance.
Under the description box, Select agent type (optional) lets you pin the type of content the agent generates before it is drafted. Leave it untouched and Cernel detects the right type from your description; pick a type to lock it in when you already know what you want (for example, choosing Single Select for a gender classifier so it isn’t mistaken for a text field).The agent types are:
  • Text: Plain text like product descriptions and meta titles
  • HTML: Rich formatted content with HTML markup
  • Single Select: Choose one value from a list (e.g., Gender: Male / Female / Unisex)
  • Multi Select: Choose multiple values from a list (e.g., Materials: Cotton, Polyester)
  • External Data: Data fetched from an external source (e.g., GTIN matching)
  • Media: Image content generated or sourced by AI
  • Picker (Preview): Instead of generating a value, it selects the best one, or several, from a set of candidates. See Picker agent below.
Pin the agent type when your description could read more than one way. “Tag the season” might generate as free text, but selecting Single Select (or Multi Select) guarantees the agent returns picks from a fixed list of options.
4

Configure the prompt

Whichever way you started, the agent’s instructions live on the sandbox’s Prompt tab, where you can edit them directly or ask the chat to rewrite them. Be specific about:
  • What to generate: the exact type of content (e.g., “a 150-word product description”)
  • Tone and style: your brand voice (e.g., “professional but approachable”)
  • What to include: specific details the output must cover
  • What to avoid: content or phrasing to exclude
Referencing your product’s attributes and dataWhether you wrote the prompt yourself or the AI generated it, type / in the prompt editor to reference the product’s own attributes and data directly in your instructions. The reference menu lets you insert:
  • Product data: all of the product’s attributes at once (All product data), or any single attribute. Every attribute is listed under the Product attributes caption inside Product data, the current category’s own attributes first (with their type shown), then the rest of your organization’s attributes, including values other agents have already generated on the same product.
  • Product title, Product images, and GTIN: shortcuts to the most-used product fields.
  • Product identifiers: a specific identifier such as SKU, GTIN, or MPN.
  • Product defining attributes: the values that make this product a distinct variant, such as its Size or Color, resolved from wherever the variant axis was introduced in the product hierarchy.
  • Data Sources: entries from your organization’s context library (brand voice, product rules, compliance guidelines).
Choosing which product data the agent getsA Product data pill in an editable prompt ends in a small arrow. Click the pill to open its options and decide exactly which of the product’s data the agent receives. Each change applies to the pill straight away, and the pill reads your choices back as a sentence, for example Product data limited to [Base data] except [Color]. Click Done to close the panel.
An agent's Prompt tab in edit mode with the Product data pill clicked, showing the Product data options panel: a Data choice of All, Base or Enriched, an Attributes section with Leave out and Only include, an Include images switch, a Variants section, and a Done button
For All product data, the panel offers:
  • Data: All, Base, or Enriched. Base data keeps an attribute only while its value comes from an integration or a CSV import. Once an agent or a person changes the value, the attribute counts as Enriched data. Pick Base when the agent should work from your supplier data only, without reading values other agents have written.
  • Attributes: Leave out the attributes you pick, or Only include the attributes you pick. Use this to keep attributes out of the agent’s input when they would distract it or don’t belong in what it writes.
  • Include images: also sends image values to the model as images. It’s off by default, since images cost more and take longer.
  • Variants: turn on Fill in from variants to fill the product’s empty values from the first variant that has one. Attributes that only the variants have aren’t added this way. To use one, add it on its own with / and turn the option on for that pill.
A pill for a single attribute offers Fill in from variants (use the product’s own value, and if it has none, the value from the first variant that has one) and the same Data choice.Read-only prompts show the sentence without the arrow. If you type a placeholder by hand with options Cernel can’t read, the pill says so, and the text is sent to the model exactly as written.Typing / followed by a space (as in “short / no emojis”) is treated as punctuation and doesn’t open the menu. If a search matches nothing, pressing Enter closes the menu and adds a new line as usual.
External Data and ID are no longer offered in the / menu. Find a specific identifier under Product Identifiers instead. Prompts that already reference External Data or ID keep their pills and keep working.
Each reference appears as a labeled pill inline with your prompt text. When the AI runs, the pill is replaced with the actual content. This keeps your prompts clean and maintainable - store your brand guide once in Tools > Data Sources and reference it wherever you need it. Update the Data Source, and every agent using it picks up the change automatically.The reference picker shows a clickable breadcrumb trail at the top of each submenu (for example, Data Sources > Models > ToV), so you can jump straight back to a parent level when you’ve dug too deep. Search inside the picker is space-tolerant: type Specifications type 1 and it matches entries named specifications_type_1.
An agent open in the Agent Sandbox on its Prompt tab, with prompt instructions that reference the product title, product data, product images and product identifiers as pills, and the agent chat on the right
Start simple. A focused 3-4 sentence prompt often outperforms a lengthy one. You can always add detail after reviewing the initial results.
5

Test in the sandbox and refine

Before saving, test your agent on real products in the Agent Sandbox. The sandbox picks five products for you when it opens. To change them, click Select products and pick up to 10, or open the arrow beside it and choose Select 5 random products for a fresh sample. Then click Run test. You’ll see the AI-generated output for each product, including the reasoning behind its choices and the source data it used.If the output isn’t quite right, use the agent chat on the right side of the screen to tell the AI what to change. For example: “Make the descriptions shorter” or “Don’t mention competitor brands.” You can reference specific products in the chat using @ to point out exactly what needs improving. The chat updates your prompt based on your feedback. Click Run test to see the change on the same products, or ask the chat to run the test for you and explain the results.Each refinement creates a new version of your prompt. Open the Changes tab to see what each change did, and use Restore this version on an earlier one to go back to it if a change didn’t improve things. Keep iterating: adjust the prompt through chat, review the new results, and repeat until you’re happy. If the agent should check its own output before it reaches your catalog, add Review Steps (an add-on module, not enabled by default) before you save. See Agent Sandbox for the full reference, including @-mentions, the detail drawer, and per-type behavior.
Test with products from different categories and price points. An agent that works well for running shoes might need adjustment for dress shoes. Catching these edge cases in the sandbox saves time later.
6

Save the agent

Once you’re satisfied with the output, save the agent. It’s now available to link to attributes on any category in your taxonomy.
You’ve created an AI agent. Next, link it to an attribute on a category in your Cernel Taxonomy to start generating content for products.

Linking an agent to a category

1

Navigate to a category

Go to your Catalog in the sidebar and select the category where you want this agent to generate content.
2

Open the attribute settings

Find the column of the attribute you want the agent to generate (or create the attribute first, see Cernel Taxonomy). Open the column’s menu from its header and choose Configure.
3

Link the agent

The Configure dialog shows the agent linked to the attribute, if there is one. Click Select agent to link one of your existing agents, or Create new agent to build one. Both continue in the Agent Sandbox: choose Use existing agent on its start screen and click Link agent, or create the new agent there. When you’re done, the agent is linked to the attribute and you return to the catalog. It will generate values for this attribute whenever you run enrichment on products in this category.
Each attribute can only have one agent linked at a time. To replace it, click Switch in the Configure dialog and pick or create another agent. Click Unlink to stop the agent from generating values for that attribute.
The agent is linked. Products in this category (and all subcategories) will use this agent when you run enrichment.

Running enrichment

1

Select products to enrich

From the product table, select the products you want to generate content for. You can select individual products or use bulk selection to pick all products in a category.
2

Choose attributes and start enrichment

Click Enrich Products. The enrichment modal shows the available attributes (those with linked agents) and how many products will be processed. Select the attributes you want to generate and click Start Enrichment.
3

Monitor the job

Enrichment runs as a background job. You can track progress from the Dashboard, which shows active jobs with their status: Pending, Running, Completed, or Failed.
The Job Details page showing a list of enriched products with columns for Product Name, Status (Review/Action), Score, and Language

Job details showing products with their enrichment status and scores

4

Review and approve results

Once the job completes, review the AI-generated content. Each product shows the generated values alongside the original data. You can approve, edit, or reject each result before it gets pushed to your connected integrations.
The Job Review page showing a product with its image and attributes on the left, and AI-generated content on the right with AI Reasoning and Source Data dropdowns, plus Review and Apply controls

The review interface: inspect AI reasoning, view source data, and approve results

You’ve successfully run enrichment and reviewed the results. Approved content is ready to sync to your connected integrations like Shopify.

Advanced configuration

The Prompt Template Library gives you pre-built agents for common content types. Each template includes a tested prompt, the right attribute type, and configuration that’s ready to use.To use a template:
  1. On the Agents page, click Create Agent to open the Agent Sandbox.
  2. Under Other options, click Use a template.
  3. Browse the library: search by name, or filter by agent type. Where the library has them, a Popular view lists a handful of hand-picked templates first.
  4. Pick a template and click Use template. The sandbox opens with the template’s configuration filled in, so you can test it on your products straight away.
  5. Give the agent a name, review the prompt, and click Create agent.
Templates are starting points, not final configurations. After creating the agent, edit the prompt to add your brand voice, reference your Data Sources, and adjust the instructions for your specific products.Recommended workflow for new users:
  1. Connect your product source (Shopify, CSV, Feed, or API)
  2. Add the attributes your categories need
  3. Create agents from templates for each attribute
  4. Test on 5-10 products and review quality
  5. Customize prompts - add your brand voice via Data Sources
  6. Scale up - enrich your full catalog or set up Automations
When reviewing generated content, you can view the Agent Reasoning, which is a breakdown of how the AI arrived at its output. This is helpful for understanding why the AI chose specific wording or values, and for refining your prompts.
Some agents can depend on the output of other agents. For example, a “Meta Description” agent might reference the output of a “Product Description” agent to create a shorter summary. Configure dependencies on the Dependencies tab in the Agent Sandbox.When you run enrichment, Cernel processes dependent agents in the correct order automatically.
Text, HTML, Single Select, and Multi Select agents let you choose the AI Provider that powers them: Google, OpenAI, or Anthropic. You’ll find it in the agent’s Advanced Settings tab. You choose the provider, and Cernel chooses which of that provider’s models runs the agent. The same provider translates the agent’s values into any language set to Translate.Model Behavior (Precise, Balanced, or Creative) is a Media agent setting. It controls how much the generated images vary, from the most consistent output to more imaginative results. Other agent types don’t have it.There is no separate setting for output length. To change how long the output is, say so in the prompt, for example “Write a 150-word product description”.
By default, when an agent enriches a variant it can draw on values the variant inherits from its parent products, so a variant with no image of its own can still be described from the parent’s image. See When I enrich a variant, does it use the parent product’s data? for how that lookup works.You can turn this off per attribute. In the attribute’s settings for a category, under the Attribute section at the top, switch off Variants inherit this from their parent. The setting is saved with the attribute, whether or not an agent is linked to it.With it on, a variant with no value of its own for the attribute shows, filters, and counts as having its parent’s value, and the attribute’s agent also reads the parent’s data. With it off, variants hold only their own value: the product table shows nothing inherited, filters with Search within variants match the variant’s own value only, Only enrich missing fields treats the variant as missing, and the agent enriches each variant from the variant’s own values only. Turn it off when your parent products hold placeholder or copied data that would otherwise leak into variant content.The category, the governing agent settings, and the variant-defining attributes (such as size or colour) are still read through the parent even when this is off. Only inherited attribute values are withheld. If the agent needs an attribute that only the parent holds, enrichment fails for that variant, and Cernel warns you of this when you switch the setting off.
Whether new values wait in Review before they sync is normally decided for each enrichment run, starting from your organization’s default. You can override that for a single attribute, so its values always wait for review (or always skip it) no matter what a run asks for.In an agent’s settings for a category, open the Review for this attribute panel and switch on Override review setting. Then choose one of:
  • Require review: new values for this attribute wait in Review until someone approves them.
  • Approve automatically: new values for this attribute skip review and are sent to your connected site.
Your organization’s default is named next to the choice, so you can see what the attribute would otherwise follow. Turning the override on starts at Require review, so nothing goes live by accident. Leave the override off and each run decides as before.Use Require review for attributes where accuracy matters most (for example a safety or compliance field), and Approve automatically for low-risk attributes you are happy to publish without a check.
External Data agents that search the web for text data (specifications, descriptions, materials sourced from manufacturer sites) have two modes. You pick the mode in the search configuration step when you build the agent.Classic mode. You define exactly which search queries to run and which sites Cernel pulls results from. Cernel searches those sources, then extracts and combines the results following your prompt. Best when you already know the canonical sources for a product type (e.g., a single supplier portal or a fixed list of vendor pages) and want predictable, repeatable searches.Agentic mode. Cernel plans and runs the searches for you, one product at a time, based on the product’s own details. You don’t write search queries, the agent reads the product and decides what to look up. You still control which domains are in scope through two lists in Website Filtering:
  • Block these domains: results from these domains are excluded
  • Restrict to these domains: when non-empty, only these domains are considered
Enter domains without the http:// or https:// prefix. Write example.com, not https://example.com/. An entry that includes the protocol prefix never matches any result: a blocked domain won’t be excluded, and a restricted list won’t return any results.
You can add sites one at a time or paste a whole list at once. Type or paste several domains separated by commas, spaces, or line breaks and Cernel adds them together. Any entry it can’t read as a domain stays in the box with a hint so you can fix it, and Clear all empties a list in one click. The same editor handles the source list in Classic mode.Use Agentic mode when products vary too much for one query template to cover them. For example, sourcing technical specs across thousands of SKUs from manufacturers you don’t enumerate up front.Product Matching appears in both modes. It tells the agent which web result actually belongs to the product in front of it. You can:
  • Match by [identifier]: pick a product identifier such as SKU, GTIN, or EAN. The agent requires that identifier to appear in the result to count as a match. Highest accuracy when you have identifiers in your catalog.
  • Match by Broad search: match on the product’s title and details instead. More flexible, but more likely to surface inaccurate results.
  • Custom match criteria (optional): add extra words or codes a result must contain. Leave blank to match on the title automatically.
What to Extract (the prompt at the bottom of the search step) tells the AI what to pull out of the matched results and how to present it. Type / to insert references to product data, Data Sources, or other attribute values, exactly like the prompt step for a Text agent.
Try Agentic mode first when your products are highly varied or when you don’t have a fixed list of vendor sites. Switch to Classic when you have a small, known set of sources and want full control over the queries.
Create an External Data agent in image mode to search the web for product images, extract them from pages, and return a curated list. Image mode uses the same Classic configuration described above for text, you define the search queries and extraction prompt yourself. (Agentic mode is for text data only.)The pipeline has three stages:
  1. Search: The agent runs web searches using queries you define (e.g., product name + brand + “product photo”)
  2. Extract: Cernel visits the search result pages and extracts images, guided by your extraction prompt
  3. Output: A deduplicated list of images is returned as the attribute value, ready for review
To set it up:
  1. Create an agent with External Data as the attribute type, then switch to Image mode
  2. Define search queries that reference product data (e.g., {product_name} {brand} product photo)
  3. Write an extraction prompt that tells the AI what images to look for (e.g., “Extract high-resolution product photos, ignore thumbnails and navigation images”)
  4. Link the agent to an External Data attribute on your taxonomy and run enrichment
Chaining with Media agents: Once you have web-sourced images, pass them to a Media agent that selects the best image, generates variations, or creates styled photography. Set up the dependency in the Media agent’s configuration so it references the External Data attribute’s output. Cernel processes agents in the correct order automatically.
Be specific in search queries - include brand name, product type, and image style (e.g., “on model”, “flat lay”). Use the extraction prompt to filter noise like navigation icons, banners, and ads.
Media agents that generate images have extra settings on the Advanced Settings tab in the Agent Sandbox. They control both how the image looks and the metadata Cernel attaches to it.
The Media agent editor's Advanced Settings tab showing Aspect Ratio, Resolution and File Format selectors, an Alt Text Instructions field, and a File Name Instructions field, each with a 'Type / for data sources' hint

The Advanced Settings tab of a Media agent: image output settings plus Alt Text and File Name instructions

Image generation modelPick the provider that generates the images. Each card shows the model it runs:
  • Google: Nano Banana 2.1. This is the default for new Media agents.
  • OpenAI: GPT Image 2.6 Sunburst.
A Media agent's Advanced Settings tab showing the Image generation model setting with a Google card (Nano Banana 2.1, selected) and an OpenAI card (GPT Image 2.5 Sunburst), followed by the Model Behavior options Precise, Balanced, and Creative

The Image generation model cards on a Media agent's Advanced Settings tab, with Model Behavior below

Choosing a card switches the agent to that card’s model. Agents you created before keep the model they already use, and the selected card names that model, until you click a card. For example, an agent set up on GPT Image 2.5 Sunburst keeps using it until you click the OpenAI card, which moves it to GPT Image 2.6 Sunburst. Model Behavior (Precise, Balanced, or Creative) only appears with the Google card, because OpenAI’s image models don’t use it.Image output
  • Aspect Ratio: the shape of the generated image (for example, 9:16 for a tall product shot).
  • Resolution: the output image quality. Leave it on Default to use the model’s native output.
  • File Format: the output file type: Default (model output), PNG, JPEG, or WebP.
Alt Text InstructionsTell the AI how to write each image’s alt text, the short description used for accessibility and SEO. For example: “Describe the product, its color, and the setting in one sentence.” Leave the field empty to use Cernel’s default alt text.File Name InstructionsTell the AI how to name the generated image file. For example: “Use the product name and color, separated by hyphens.” Descriptive file names help with image SEO and keep your media library organized. Leave the field empty to use Cernel’s default naming.Both instruction fields work like the main prompt editor: type / to insert references to product data, Data Sources, or other attribute values, so the alt text and file name can pull in details like the product title or color automatically.
Keep alt text instructions to a single, factual sentence. Search engines and screen readers both favor concise, descriptive alt text over keyword-stuffed phrases.
Preview feature. Background recoloring is in early access. Behavior may change, so review the results before relying on them for production imagery.
Media agents that generate images can replace the background of every image they produce with a single, exact color. This is ideal for product and studio photos that need a clean, consistent backdrop, for example, when every image tile on a product page has to look identical. Lifestyle photos with real surroundings won’t look natural against a solid background, so leave recoloring off for those.To turn it on, open the Media agent, go to the Advanced Settings tab, and switch on Recolor background. Three controls appear:
  • Background color: the solid color the background becomes, entered as a hex value (for example, #FFFFFF).
  • Mode: how Cernel treats shadows behind the product:
    • Replace background: best for product shots; removes shadows as well.
    • Preserve shadows: best for studio shots of models; keeps the natural shadows.
    • Hybrid (clean background, keep shadows): keeps only the clearest shadows near the product or model and recolors everything else to your exact color. Choose it when the background must match a precise color but you still want some depth.
  • Processing engine: how firmly Cernel trims the edges around the subject:
    • Standard (recommended): keeps the most natural edges; the right choice for most images.
    • Alternative: trims edges more firmly. Pick this if you notice a faint halo around the product or model, though it can give harder edges around fine details like hair.
The Recolor background section of a Media agent's Advanced Settings tab, showing a Background color hex field, a Mode selector with Replace background / Preserve shadows / Hybrid options, and a Processing engine selector with Standard and Alternative options

The Recolor background section in a Media agent's Advanced Settings: the Background color, Mode, and Processing engine controls

In Settings > Global Prompts, you can set one prompt per attribute type (Text, HTML, Single Select, Multi Select). It’s added to every agent of that type that has Include default system instructions switched on in Advanced Settings (on by default). Use it only for rules that must hold for every value of the type, like “never use superlatives” for all Text attributes. Instructions for one attribute belong in that agent’s own prompt. Turn the switch off for an agent that should ignore the Global Prompt. See Global Prompts.
Text and HTML agents have a Languages tab with one instruction box per enabled language, so you can give a single agent language-specific guidance without changing your organization-wide language settings. A toggle keeps or replaces the organization’s language instructions for that agent. Other agent types don’t have the tab, because their values come from option labels or source data rather than model-written text. See Give an agent its own language instructions for the full walkthrough.
Re-running enrichment on products that already have approved values will generate new values. The previous approved content will be overwritten if you approve the new results. Review carefully before approving regenerated content.

Picker agent

Preview feature. The Picker agent is in early access, and its behavior may change before it’s generally available.
Most agents generate a new value for an attribute. A Picker agent does something different: it selects the best value, or values, from a set of candidates. Think of choosing the single cleanest packshot, the strongest tagline, or the top few options from a longer list. Reach for it when the right answer already exists among several choices and you want the AI to pick rather than write.

How a Picker agent works

A Picker runs as a two-phase pipeline: first it gathers candidates, then it picks from them. Phase 1: Find candidates. A search prompt lists the candidates to choose from. Reference product data with {{placeholders}} such as {{images}} or {{title}} (for example, “List all available product images: ”). By default, candidates are read straight from the prompt. You can optionally widen the net with Data access:
  • Products: let the agent search your product catalog (useful for “shop-a-look” or look-alike picks). Pin a category and filters to focus the search.
  • Data Sources: let the agent browse your organization’s reusable content to source candidates.
Phase 2: Select. The Strategy controls how the agent picks from the candidates:
  • AI ranks them: a second AI pass scores the candidates against a Rubric you write (for example, “Pick the cleanest, frontal packshot with a plain background.”). A rubric is required for this strategy.
  • Random pick: skip the AI and choose at random. It’s fast, free, and reproducible: set a Seed and the same candidates always produce the same pick (handy for tests). Leave the seed blank for a fresh pick each run.

Output

In the Output section, choose what the agent produces:
  • Values: whether the agent selects a Single value or Multiple values.
  • What it picks: whether the candidates are Images or Text.
Candidates are sampled without replacement, so the same value is never picked twice. For Multiple values, a How many? control sets how the count is enforced. It decides both how many values the agent keeps and whether a thin set of candidates passes or fails the enrichment:
  • All that fit: every candidate the agent finds is kept, in alphabetical order. No selection step runs.
  • At most N: the agent keeps up to N values, fewer when only some candidates fit, and none (saving no value) when none of them do. Choose this when a partial result is still useful.
  • At least N: the agent must find at least N values. If it finds fewer, nothing is saved and the agent reports an error.
  • Exactly N: the agent must find exactly N values. If it finds fewer, nothing is saved and the agent reports an error.
  • Between N and M: the agent keeps between N and M values. If it finds fewer than N, nothing is saved and the agent reports an error.
So At most and All that fit never fail on a thin candidate set: they save whatever fits, down to no value at all. At least, Exactly, and Between enforce a floor and fail the run loudly when the agent cannot reach it. The step heading and config summary reflect your choice, for example Up to 3 or Exactly 3.

What a text Picker outputs

When a Picker outputs Text and has Products data access turned on, it picks whole products from your catalog, so you can choose what each pick actually carries. An Each pick outputs setting appears with three options:
  • Product title (the default): every pick is the title of the product the agent chose.
  • An attribute: every pick is the chosen attribute’s value on the product the agent picked. Pick the attribute with the selector that appears. Products without a value for that attribute are skipped.
  • What the prompt asks for: the agent reads your search prompt to decide what to output, for example “pick a product’s material”. It falls back to the title when the prompt names nothing.
An Image Picker always outputs the image, so this setting applies to text output only.

Creating a Picker agent

1

Choose the Picker type

Start a new agent (see Creating an agent) and select the Picker type. Cernel opens the Picker Configuration settings.
2

Set the output

Choose Single value or Multiple values, and whether the candidates are Images or Text. For multiple values, set How many? (All that fit, At most, At least, Exactly, or Between) to control how many values are kept and whether a thin candidate set fails the run. See Output.
3

Write Phase 1: find candidates

Write the search prompt that lists the candidates, referencing product data with {{placeholders}}. Turn on Data access to Products or Data Sources if you want the agent to look beyond the prompt.For a Text Picker with Products access, set Each pick outputs to control what each pick carries: the product title, a named attribute, or whatever the search prompt asks for. See What a text Picker outputs.
4

Set Phase 2: select

Pick a Strategy. For AI ranks them, write a Rubric describing what makes a good pick. For Random pick, optionally set a Seed for reproducible results.
5

Save and link it

Save the agent and link it to an attribute on a category, just like any other agent.
Try a Picker in the agent sandbox before linking it to a category, so you can see which candidate it selects and refine your rubric.

Frequently asked questions

Just edit the agent. Go to Agents, find the agent, and update its prompt. Because the agent is linked to categories (not copied), every category using that agent will use the updated prompt the next time you run enrichment. You don’t need to update each category separately.
Check three things:
  1. Your agent prompt: make sure it includes specific style and tone instructions, not just what to generate. Open the agent, run the sandbox, and use the agent chat to tell the AI exactly what to adjust.
  2. Data Sources: reference a Data Source containing your brand guidelines in the prompt. Type / in the prompt editor to insert one.
  3. Global Prompts: check Settings > Global Prompts for any system-level instructions that might conflict with your agent’s style.
Use the sandbox to test changes on a few products before running enrichment on your full catalog.
A Picker’s How many? setting controls this. At least, Exactly, and Between enforce a floor: if the agent finds fewer values than you asked for, nothing is saved and the run reports an error. Switch to At most or All that fit when a partial result is still useful. Those keep however many candidates fit, down to none (which saves no value), instead of failing the enrichment. See Output.
No. Each agent is configured for a specific attribute type (Text, HTML, Single Select, Picker, etc.). If you need similar content in different formats, create separate agents (one for Text and one for HTML, for example) with similar prompts.
Yes. On the Agents page, click Create Agent, then choose Copy from existing on the start screen and pick the agent to copy. The Agent Sandbox opens a copy with the same settings, named after the original with “Copy” at the end. The original stays as it is, and the copy is only created when you click Create agent, so you can change the prompt and test it first.
You can’t delete an agent while it’s in use. When you click Delete agent on an agent that’s linked to categories, the Agent In Use dialog lists those categories instead, so you can unlink the agent from each of them first. An agent that isn’t linked anywhere is deleted once you type its name to confirm and click Delete Permanently. Previously generated content is not deleted.
Yes. Open the Media agent, go to the Advanced Settings tab, and fill in Alt Text Instructions and File Name Instructions. These tell the AI how to describe and name each generated image. You can reference product data (like the title or color) by typing / in either field. Leave a field empty to use Cernel’s default. See Media agents: image output, alt text, and file names.
You can enrich an entire category at once, regardless of size. Cernel processes enrichment as background jobs, so large batches run without blocking the rest of the platform. You can monitor progress from the Dashboard.
Yes. By default, enrichment results go through a review step. You can enable Approve Before Push in Settings > Platform to require explicit approval before any content syncs to your connected integrations. This gives you full control over what gets published.

What’s next

Cernel Taxonomy

Set up your category structure and link agents to attributes so you can run enrichment.

Connecting Shopify

Connect your Shopify store to import products and push enriched content back.