Prompt Steps

AI-powered steps that can think, research, write, and interact with tools.

Overview

Prompt steps are the heart of most workflows. They use AI to process information, generate content, make decisions, and interact with external tools and services. Each prompt step runs a conversation with the AI model to accomplish its task.

Key capability: Prompt steps can use tools to send emails, search the web, access files, interact with integrations, and much more.

Configuration

Each prompt step has the following configuration options:

Property	Type	Description
name	string	Display name for the step
prompt	string	Instructions for what the AI should do
tools	string[] \| ToolSet	Tools available to this step
modelId	string?	AI model override (optional)
needsApproval	boolean?	Require approval before send actions
disabled	boolean?	Skip this step when running

Writing Effective Prompts

The prompt field tells the AI what to do. Write clear, specific instructions.

Be specific about the task

Research the latest trends in [topic] and summarize
the top 5 findings. Include sources for each point.

Reference inputs from previous steps

Using the research from the previous step, write a
professional email to share these findings with
the team. Keep it concise and actionable.

Specify output format when needed

Analyze the provided data and create a summary table
with columns: Category, Count, Trend (up/down/stable).

Available Models

By default, prompt steps use Claude 4.5 Sonnet. You can override the model for specific needs.

Model	Best For
anthropic:claude-4.5-sonnet	Default. Balanced speed and quality
anthropic:claude-4-sonnet	Previous generation, cost-effective
openai:gpt-5.2	Alternative for specific use cases
google:gemini-3-flash-preview	Fast responses, lower cost
google:gemini-3-pro-preview	High quality, complex reasoning

Tool Access

Prompt steps can use tools to interact with external services. Tools are explicitly defined per step - there is no implicit access.

Common tool categories:

Communication: Send emails, SMS, Slack messages
Research: Web search, browse pages
Files: Read, write, upload files
Data: Access spreadsheets, databases
Integrations: CRM, project management, etc.

Guardrails (needsApproval)

Enable guardrails to add a safety check before the AI executes "send" actions like emails or messages. This is separate from Approval Steps.

How guardrails work:

Prompt step runs and prepares to send a message
Workflow pauses before executing the send tool
You review the exact content that will be sent
Approve to send, or reject to skip the action
Workflow continues to next step

Learn more in the Approvals Guide.

Analytics & Monitoring

Each prompt step tracks detailed analytics for monitoring and cost management.

Metric	Description
inputTokens	Tokens sent to the model
outputTokens	Tokens generated by the model
cachedInputTokens	Tokens served from cache (lower cost)
cost	Total cost in USD
toolCallCount	Number of tools used
durationMs	Execution time in milliseconds

Message History

Prompt steps capture the full conversation for debugging and audit trails. This includes user messages, assistant responses, and tool calls with their results.

{
  "role": "user",
  "content": "Research AI trends..."
}
{
  "role": "assistant",
  "content": "I'll search for the latest...",
  "tool_calls": [{ "name": "webSearch", "input": {...} }]
}
{
  "role": "tool",
  "content": "Search results: ..."
}

Example Configuration

{
  "id": "step-1",
  "type": "prompt",
  "name": "Research Topic",
  "prompt": "Research the provided topic and summarize key findings. Include at least 3 sources.",
  "tools": ["webSearch", "browsePage"],
  "modelId": "anthropic:claude-4.5-sonnet",
  "needsApproval": false
}

Best Practices

Keep each prompt step focused on a single task
Enable guardrails for steps that send external communications
Only include tools that are actually needed for the step
Use clear, specific language in prompts
Reference outputs from previous steps explicitly
Consider model costs for high-volume workflows

Approvals Guide - Deep dive into approval patterns and guardrails
Approval Steps - Dedicated content review steps
Building Workflows - Overall workflow design