First Workflow

Use a Workflow to complete a multi-step task: input topic → generate outline → expand body → output Markdown. About 10 minutes.

What You'll Accomplish

A 4-node Workflow:

[Start: input topic] → [LLM: generate outline] → [LLM: expand body] → [Output: Markdown]

Steps

1 · Create the Workflow (1 minute)

1. Workspace → Apps · Workflow → New → Blank Workflow
2. Name it: Article Generator
3. Open the visual editor

2 · Configure the start node (1 minute)

The start node is added automatically. Click it and define inputs:

3 · Add an LLM node · Generate Outline (2 minutes)

Drag an LLM node from the panel and connect it to the start node. Configure:

• Node name: Generate Outline
• Model: GPT-4 / Claude (your choice)
• Temperature: 0.7
• Prompt:

  You are a senior content writer. Based on the topic "{{topic}}", generate an outline for an article of {{length}} words in {{tone}} style.
  
  Output JSON:
  {
    "title": "...",
    "sections": [
      {"heading": "...", "key_points": ["...", "..."]}
    ]
  }

• Output variable: outline (type: JSON)

4 · Add a second LLM node · Expand Body (2 minutes)

Drag a second LLM node and connect it after the first. Configure:

• Node name: Expand Body
• Model: same as above (can differ)
• Temperature: 0.7
• Prompt:

  Based on the outline below, write a complete article of about {{length}} words in {{tone}} style.
  Output in Markdown, no metadata.
  
  Outline:
  {{outline}}

• Output variable: article (type: string)

5 · Add an output node (1 minute)

Drag an Output node. Configure:

• Output field: article (bind directly to the previous node's article)

6 · Test (2 minutes)

Top right of the editor → Trial run → input:

{
  "topic": "Common pitfalls of RAG in enterprise rollout",
  "tone": "deep technical",
  "length": 1500
}

Watch the execution flow. Click any node to see inputs / outputs / latency / token counts.

7 · Publish (1 minute)

• Save version → note: v1 · initial
• Publish to Workbench
• Set the trigger:
  • Manual (default) — users trigger from the Workbench
  • Scheduled — configure Cron in Schedule
  • API — for external systems; see Workflow API

Use Traces to Understand Multi-Node Coordination

Workspace → Workspace management · Observability · Traces — find the run record:

Workflow.run [3.2s · 4 nodes]
├─ start [0.01s] → input: {...}
├─ Generate Outline [1.4s] → tokens in/out: 220/580 · cost: 0.012 Credit
├─ Expand Body [1.7s] → tokens in/out: 800/1500 · cost: 0.038 Credit
└─ end [0.02s] → output: {...}

Every step's cost and latency can be attributed. This is the core value of Workflow over calling LLMs directly.

Going Further

→ Workflow full capability

Next Steps

• Want a conversational app → First Agent
• Want real-world scenarios → Customer service bot · Marketing automation
• Want external triggers → API · Workflow