AI Generate in the builder header takes a description of the campaign and draws the flow. It runs on the chat-slot model from AI Models, and the tokens go against your own provider key.

​
What happens when you generate

1

The prompt is assembled

EDGR gives the model the catalog of available nodes with their fields, the customer fields you can reference with {{customer.*}}, and the workflow currently on the canvas.
2

The model returns a graph

It answers with JSON describing nodes and arrows against a fixed schema.
3

The graph is compiled

A graph with a cycle, a missing field or a bad node reference is rejected and nothing is applied.
4

The canvas is replaced

Every existing node is deleted and the new ones take their place. Auto-layout in the header tidies the arrangement.
Generating wipes the nodes already on the canvas. Duplicate the campaign first if there is work you want to keep.
The model is capped at 8192 output tokens, which is roughly fifteen substantial nodes. Beyond that, build the rest by hand or generate in stages.

​
Writing the prompt

State the goal, then number the steps, then say what each branch does. The model follows an explicit sequence far better than it infers one.

Name the goal

“3-step cold outreach for fintech leads, aiming at a 5% reply rate.”

Number the steps

One step per line, in order, including the waits.

Spell out the branches

“If they reply, hand off. If they clicked but did not reply, wait 5 days then follow up. If no opens after 7 days, cooldown.”

Point at variables

“Personalize the subject by industry. Mention a use case for {{customer.country}}.”

​
Example prompts

3-step cold email for fintech leads in VN.
Step 1: intro mail, personalize by industry, subject under 60 chars.
Wait 3 days.
Step 2: follow-up if not opened, mention one use case from the knowledge base.
Wait 5 days.
Step 3: exit cooldown if no click. If reply at any point, handoff to sales.

Re-engagement for customers with no contact in 30 days.
Start: short email, "Still interested?"
Wait 7 days.
If opened and clicked: adjust-score +0.2 and add tag "warm-reengaged".
If not opened: suppress with reason manual.

Onboarding for customers who just closed a won deal.
Day 1: welcome email and product overview.
Day 3: tutorial video link.
Day 7: case study.
Day 14: feedback survey over LinkedIn DM.
If they answer the survey, assign owner to customer success.

​
Steering the generator

The instructions the model receives come from a prompt template named flow-generator. Create a template with that name under Templates and it overrides the default — which is how you push the generator toward short flows, email-only designs, or your own house style.

​
After generating

The output is a skeleton. Before activating:
  1. Read the subject and body on every messaging node and rewrite them in your own voice.
  2. Swap inline bodies for template references wherever the content is reused.
  3. Replace the placeholder credentials with real ones.
  4. Adjust the Wait durations for the audience’s timezone.
  5. Clear the compile errors.
  6. Run it against one or two customers.