AI Workflow Assistant
The AI assistant generates complete workflows from plain English descriptions. It creates nodes, edges, context, template expressions, and configuration, then applies the result directly to the visual builder canvas.
Where to find it
The assistant is available in two places:
- Builder sidebar: A resizable panel on the right side of the workflow canvas at /dashboard. Click the assistant button to open it.
- Standalone page: A full-page chat interface at /dashboard/assistant for longer conversations.
Both interfaces generate the same output. The builder sidebar is convenient for quick iterations since you see the canvas update in real time.
What it generates
When you describe a workflow, the assistant returns:
- Nodes with HTTP method, URL, headers, and body configured
- Edges defining the execution order between nodes
- Context with input variables your workflow expects
- Template expressions for passing data between nodes (e.g.
{{ results.get_order.data.id }}) - Pre-built node configuration with schema-driven parameters when using hosted integrations
The generated workflow is applied to the canvas automatically. You can then review each node's configuration, adjust parameters, and add your API keys before running.
Write effective prompts
The assistant performs best when you name specific services and describe the sequence of operations. Vague prompts produce vague workflows.
Prompts that work well:
- "Use the SendGrid pre-built node to send a welcome email. Use an LLM to generate the email body about our product."
- "Fetch orders from my API (docs at api.mycompany.com/docs), then send a summary to Resend."
- A multi-step prompt like:
Detailed prompt example
1. Scrape https://www.firecrawl.dev/pricing using Firecrawl
2. Use LLM to:
- Generate a script advertising Firecrawl
- Analyze and suggest best pricing tier for startups
3. Generate video using RunwayML with the advertising script
4. Send email via SendGrid containing:
- Pricing recommendation for startups
- Generated video URL
Prompts that do not work well:
| Prompt | Problem |
|---|---|
| "Send emails when users sign up" | Does not specify which email service to use |
| "Connect to my company's API" | Does not provide a documentation URL |
| "Process payments" | Does not name a payment provider or endpoint |
Use pre-built integrations
When you mention a pre-built integration by name, the assistant fetches its OpenAPI specification and configures the node with the correct URL, method, headers, and request body schema. It helps to include the phrase "pre-built node" so the assistant knows to look for a hosted integration rather than guessing a raw HTTP endpoint.
Example: "Use the SendGrid pre-built node to send an order confirmation email" produces a node with SendGrid's API schema pre-filled.
See Pre-built Integrations for the full list of available services.
Connect external APIs
For APIs not in the pre-built library, provide a documentation URL. The assistant reads the API schema and generates correct request parameters.
Example: "Call the Stripe charges API to create a charge. Here are the docs: https://stripe.com/docs/api/charges/create"
Without a documentation URL, the assistant guesses at the API structure. This often produces incorrect request shapes, missing headers, or wrong HTTP methods.
Limitations
- The assistant cannot access private or authenticated API documentation. If your API docs require login, paste the relevant request/response schema into the chat instead.
- You must provide your own API keys. The assistant does not store or manage credentials. Add them in the node configuration form or use secrets.
- Generated workflows are a starting point. Review the configuration of each node before running, especially for request bodies and authentication headers.