Skip to main content

Using prompts in AI Workflows

Learn how to write prompts and templates for AI Workflows to improve extraction, enrichment, translation and workflow output quality.

Prompts control how AI behaves inside AI Workflows.

They define:

  • what the AI should do

  • which data should be used

  • how output should be structured

  • which rules the AI must follow

Prompts are one of the most important parts of workflow configuration because they directly influence:

  • output quality

  • extraction accuracy

  • SEO performance

  • translation consistency

  • moderation workload

Well written prompts create more reliable and scalable automation pipelines.

Where prompts are used

Prompts can be used in multiple workflow actions.

Examples:

  • Attribute Extraction

  • Content Enrichment

  • Translation

  • Category Mapping

Each action uses prompts differently depending on the workflow goal.

Prompts vs templates

A prompt is the instruction given to the AI.

Example:

Generate an SEO optimized shopping description using a professional ecommerce tone focused on pet nutrition.

A template is a reusable prompt containing dynamic variables.

Example:

Product Name: {{name}}

Description: {{description}}

Templates allow workflows to automatically generate contextual prompts for every product.

Using variables inside prompts

Templates can include dynamic product variables.

Examples:

  • {{name}}

  • {{sku}}

  • {{ean}}

  • {{description}}

These variables are automatically replaced with product data during workflow execution.

Example template:

Product Name: {{name}}

EAN: {{ean}}

Description: {{description}}

This allows prompts to scale across large catalogs automatically.

Recommended prompt structure

Strong prompts usually contain:

  • Goal

  • Input data

  • Rules

  • Output requirements

  • Fallback behavior

Example structure:

Goal

Extract product attributes from the product description.

Input

Product Name: {{name}}

Description: {{description}}

Rules

  • Only extract explicitly mentioned values

  • Do not guess missing information

  • Normalize attribute values

  • Return empty values when data is missing

Output

Return:

  • Flavor

  • Lifecycle

This structure creates more predictable AI behavior.

Example Attribute Extraction prompt

Example:

You are a highly precise product attribute extraction engine.

Product Name: {{name}}

EAN: {{ean}}

Description: {{description}}

Extract the following attributes:

  • Flavor

  • Lifecycle

Rules:

  • Only extract explicitly mentioned information

  • Do not guess missing values

  • Normalize values into consistent terminology

  • Return empty values if information is missing

This prompt focuses the AI on structured extraction instead of content generation.

Example Content Enrichment prompt

Example:

Generate an SEO optimized shopping description using a professional ecommerce tone focused on pet nutrition.

Requirements:

  • Maximum 500 characters

  • Use natural ecommerce language

  • Highlight nutritional benefits

  • Avoid keyword stuffing

  • Write in Dutch

This creates more controlled enrichment output.

Example Translation prompt

Example:

Translate the Shopping Description to German using natural ecommerce language suitable for online product pages.

Requirements:

  • Preserve formatting

  • Keep product names untranslated

  • Use professional ecommerce terminology

  • Avoid literal translations

This improves localization quality significantly.

Using normalization rules

Prompts can contain normalization logic to improve consistency.

Example:

  • “rundvlees” → “Rund”

  • “volwassen” → “Adult”

  • “krokante brokken” → “Droogvoer”

Normalization helps improve:

  • catalog consistency

  • filtering

  • layered navigation

  • marketplace compatibility

Prompt strictness

Good prompts clearly define limitations.

Examples:

  • Do not guess missing values

  • Only use verified information

  • Return empty values if data is unavailable

  • Avoid generating unsupported claims

Strict prompts reduce hallucinations and incorrect output.

Using prompts in workflow chains

Prompts become more powerful when workflows combine multiple actions.

Example workflow:

Attribute Extraction → Content Enrichment → Translation

In this workflow:

  • extracted attributes improve enrichment quality

  • enriched content improves translation quality

This creates more contextual and reliable output across the workflow pipeline.

Prompt testing

Before running workflows on large catalogs:

  • test prompts on sample products

  • review generated output

  • inspect confidence scores

  • validate formatting

  • review AI reasoning

Testing helps prevent large scale workflow issues.

Prompt best practices

Be specific

Weak prompt:

Generate a product description.

Better prompt:

Generate a Google Shopping optimized product description using a professional ecommerce tone focused on pet nutrition.

Specific prompts create more reliable output.

Use category specific prompts

Different categories often require different prompt strategies.

Examples:

  • Fashion

  • Electronics

  • Pet food

  • Supplements

Category specific prompts improve output quality significantly.

Avoid overly broad instructions

Large generic prompts often create:

  • vague descriptions

  • inconsistent formatting

  • unstable output

Focused prompts usually perform better.

Improve prompts continuously

Prompt optimization is an ongoing process.

You may improve prompts to:

  • reduce moderation workload

  • improve SEO quality

  • improve extraction accuracy

  • increase translation consistency

  • create more structured output

Example workflow using prompts

Example:

Trigger selects products missing descriptions.

Attribute Extraction prompt extracts:

  • Flavor

  • Lifecycle

Content Enrichment prompt generates:

  • Shopping descriptions

  • SEO titles

Translation prompt localizes content into German.

Moderation reviews output before synchronization.

This creates a scalable AI driven enrichment pipeline.

Did this answer your question?