The Stage instructions field is what the agent runs while the lead is in that stage. There are two writing formats, and choosing the right one makes more difference than any model or personality tweak: one guides the lead through an ordered sequence, the other reacts to whatever topic the lead brings up. This article shows both, with templates to copy.
Which format to use
| Step-based stage | Topic-based stage | |
| Flow | Linear and ordered: Step 1 → 2 → 3 | No fixed flow, reacts to the message |
| When to use | The stage needs to take the lead through a sequence | The lead may arrive with any topic |
| Structure | Numbered steps | Core behavior, by topic |
| Typical examples | Reception, Screening, Data collection, Scheduling | Meeting scheduled, Qualified with ongoing support, Post-sale |
When in doubt: if you can draw the stage as “first this, then that”, use steps. If the stage is a state where the lead stays and talks freely, use topics.
Format 1: step-based stage
The instruction has three blocks.
General rules — the behavior that applies to the whole stage: what to do if the lead answers several steps at once, goes off topic, or breaks the order.
Steps — the sequence. Each step has three parts:
-
Rules — when this step should run.
-
Goal — what the agent does: send a message, run a skill, go to another step or another stage.
-
Fallback — what to do when the lead's answer is unclear or a skill failed.
Templates — ready-made, named messages that the steps reuse, instead of repeating the text inside each step.
Template to copy
You must run the steps below according to the order, the conditions set, and the lead's answers. Interpret the lead's answer to trigger the right steps and skills in the specified order.
### General rules
- If the lead answers several steps in a single message, move forward as far as possible.
- Never run a step out of order.
- If the lead takes the conversation off topic, answer based on your knowledge and repeat the last pending question.
### Steps
#### Step 1: "[step name]"
Rules:
- [when this step should run]
Goal:
- If the lead answers positively, move to Step 2.
- If the lead answers negatively, move to Step 3.
Fallback:
- If the answer is unclear, say you did not understand and repeat the last pending question.
#### Step 2: "[step name]"
Rules:
- Run after the lead answers Step 1.
Goal:
- Send the message "[template name]".
Fallback:
- If the answer is unclear, say you did not understand and repeat the last pending question.
### Templates
#### Template: "[template name]"
Message: "[exact text the agent sends]"Format 2: topic-based stage
Here there is no linear flow: each message from the lead can trigger any topic. The agent identifies the topic from context and replies.
The instruction has three parts:
General rules and reusable information — the top declares which topics the stage can handle and gathers the information several topics will use (for example, the details of the scheduled meeting, fetched by lookup skills).
Core behavior — one block per topic, each with:
-
Triggers when — the signals that make the agent enter this topic. List real phrases the lead writes, not abstract descriptions. This is what keeps the agent from mixing up two topics.
-
Goal — what to do: search the knowledge base, send a message, run a skill, or transfer the conversation. A topic that needs several turns can have numbered steps inside its own goal.
Catch-all topic — always include a final topic such as “Other topics” for anything that does not fit the previous ones. Without it, the agent has no way out.
Template to copy
Your goal is to react to what the lead asks for, understanding the context of the current message and the previous ones.
There is no linear flow. Each message from the lead can trigger any topic. Topics that need several turns have numbered steps inside the goal.
#### General rules
- If the lead answers several steps of a topic in a single message, move forward as far as possible.
- Within a topic, never run a step out of order.
- If the lead takes the conversation off topic, answer based on your knowledge and go back to the pending question.
#### General information
- [data reusable by the topics, including what comes from lookup skills]
## Core behavior
### 1) [Topic name]
Triggers when: [situation]. Signals: "[phrase the lead writes]", "[another phrase]", "[another phrase]".
Goal:
- [what the agent does: reply, look up, run a skill, transfer]
### 2) [Topic name]
Triggers when: [situation]. Signals: "[phrase]", "[phrase]".
Goal:
#### Step 1: "[step name]"
Rules:
- [condition]
Goal:
- [action]
Fallback:
- If the answer is unclear, say you did not understand and go back to the pending question.
### 3) Other topics
Triggers when: the lead brings up any topic that does not fit the previous ones.
Goal:
- [how to reply and where to take the conversation]
## Templates
#### Template: "[template name]"
Message: "[exact text]"Editor features
Inside the instructions, type # to insert platform features. They appear as colored tags in the middle of the text, and you write normally around them:
| Feature | What it is for |
|---|---|
| Another stage | Route the lead: “…and go to the #Scheduling stage” |
| Lookup skill | Fetch a piece of data and use it in the message: “…use #Look up: contact name” |
| Update skill | Save a value to a contact or conversation field |
| Conversation assignment | Transfer to a team or agent |
| Message template | Send an approved WhatsApp template |
| Media | Send an image, audio, or document configured in the agent |
Two things worth remembering:
-
Named templates are not WhatsApp templates. In the Templates block of the instruction, you name regular messages to reuse within the stage. A Meta-approved HSM template is a different feature, inserted from the
#menu. -
Save the stage. While you edit, the Stage not saved notice appears in the footer. Changes only take effect after you click Save stage — switching stages before that asks you to save or discard.
Stage criterion
Right below Stage instructions, a second field may appear: Stage criterion. It only exists when the funnel has Enable stage update with agent turned on in the Agent tab, the option that lets the AI keep updating the stage after a human agent takes over the conversation.
The two fields do not do the same thing:
| Stage instructions | Stage criterion | |
| What it is for | What the agent does and says in the stage | When the conversation belongs to this stage |
| Who reads it | The agent that talks to the lead | The AI that classifies the stage while the human handles the conversation |
| Format | Steps or topics, can be long | One short sentence, up to 500 characters |
Write the criterion as an objective rule, anchored in a fact said in the conversation, and also say what does not count:
“If the person has already paid, move them to Paid. A question about price or a discount request does not count as payment.”
What the AI does with this: if no criterion is clearly met, it keeps the current stage instead of guessing. Conversation tone decides nothing — an objection or hesitation from the lead, on its own, does not meet any criterion.
Main stages do not use this field to be chosen: they follow the rule of moving forward without going back, with or without a criterion. Behavior details in How to set up a Smart Funnel.
Run the prompt automatically on entry
In Advanced settings, at the bottom of the stage editor, the Run this stage's prompt automatically on entry toggle makes the agent run the instruction the moment the lead reaches the stage, without waiting for a new message.
This is what you want in stages where the agent needs to take the initiative — after qualifying, for example, sending the time proposal right away instead of staying silent. Behavior details and limits in How to set up a Smart Funnel.
Before saving, review
-
Does every step have a Fallback?
-
Does the topic-based stage have the Other topics topic?
-
Are the trigger signals for each topic real phrases that do not overlap?
-
Does every exit lead somewhere — another step, another topic, or another stage?
-
Is the order of the steps correct?