Agent Step Prompts

Last Updated: October 25, 2026

Agent step prompts for Workflow are natural language instructions that help you analyze data and generate or rewrite content based on data from previous workflow steps. Agent steps can return a text response, structured data, or a file. For file types and limits refer to Agent File Output Reference. Keep the output type in mind when you write prompts to optimize output and ensure workflow continuation.

This feature is in beta. Capabilities, configuration, and availability may change as development continues. Contact your Vasion representative with feedback to help shape this feature.

Operation Prompts

The Agent step interprets natural language to analyze data and generate text, structured data, or file output. The following prompts and examples demonstrate different approaches based on your needs.

Analysis Prompts

Use these prompts to analyze data from previous workflow steps.

Simple Analysis

Review <data source>, and identify <what to look for>.

The following are examples:

  • Review this invoice, and flag any line items that exceed $5,000.
  • Check this document for missing required fields, and list what is absent.
  • Analyze this claim, and list any common denial triggers.
Scored Assessment

Evaluate <data source> against <criteria>, and provide a <score or rating> with reasoning.

The following are examples:

  • Evaluate this contract against our standard terms checklist, and provide a completeness score from 1 to 10 with a list of missing clauses.
  • Assess this support request for high, medium, or low urgency, and explain your reasoning.
  • Review this app for completeness, and rate each section as complete, partial, or missing.

Generation Prompts

Use these prompts to produce new content based on workflow data.

Text Generation

Using <data from previous steps>, generate <output>.

The following are examples:

  • Using the patient data and diagnosis codes from the previous steps, generate a letter of medical necessity for prior authorization.
  • Using this form submission, draft a personalized acknowledgment email with the submitter's name and expected timeline.
  • Using these extracted invoice fields, generate a summary suitable for the approver.
Rewriting and Formatting
Rewrite <input> to meet <standard or format>.

The following are examples:

  • Rewrite these free-text notes into standardized language that meets our compliance formatting requirements.
  • Reformat this extracted data into a clean summary with these labeled fields: Name, Date, Amount, and Status.
Structured Data
Using <data from previous steps>, return <fields> as structured data.

The following are examples:

  • Review this invoice, and return the vendor name, invoice date, and total amount as structured data.
  • Using this form submission, return the requester name, department, and request type as structured data.
File Generation

Using <data from previous steps>, generate a <file type> file that contains <content>. The agent must have the Code Interpreter tool.

The following are examples:

  • Using these extracted invoice fields, generate an Excel spreadsheet that lists each line item, quantity, and amount.
  • Using this form submission, generate a PDF summary suitable for the approver.

  • Using these extracted invoice fields, generate a JSON file named invoice-data.json.

Prompt Approaches

Choose your prompting approach based on how well-defined the task is and the output you need.

Direct Instruction

Use clear, specific prompts when you know exactly which output you need. This approach works best for well-defined tasks.

You can do the following:

  • Tell the agent exactly what to look for and how to format the output.
  • Specify the structure, such as "Respond with: <field 1>, <field 2>, <field 3>."
  • Include the criteria or rules the agent should apply.

Contextual Analysis

Use open-ended prompts when you want the agent to interpret and reason about the data. This approach works best for review and assessment tasks.

You can do the following:

  • Let the agent identify issues, patterns, or anomalies.
  • Ask for reasoning alongside conclusions.
  • Refine with follow-up instructions in subsequent Agent steps.

Recommendations

Follow these recommendations to get the best results from the Agent step:

  • Define the output structure: Tell the agent how to format the response. To use values in later steps, ask for structured data.

  • Request exact outputs: If a later step requires an exact output, tell the agent to return only that output. If the agent returns something else, the workflow does not continue.

  • Reference previous step data explicitly: Mention which data is coming from the prior step so that the agent knows which inputs to expect.
  • Keep instructions focused on one task: One Agent step should do one cognitive job. Chain multiple Agent steps for multistage analysis rather than combining everything into one prompt.
  • Test with real workflow data: Prompt quality depends on the data in the workflow. Test with actual documents and form submissions.
  • Send output to email: Add an Email step after the Agent step. Include the agentResult variable, and select the agent name to attach the file.. For more details refer to Add an Email Step.

Next Steps

Refer to the following: