Skip to main content

​
Frequent mistakes

These are the mistakes we see repeated most often when reviewing bots — most are easy to avoid once you know they exist.
If you write {variable} in the Initial Message or in the Instructions and that variable never reached the context — neither through the campaign file nor the Initial Code — it stays empty during the conversation instead of raising a visible error. Before publishing, check that every variable you use is actually mapped or initialized. See Available variables.
Labels are predefined by Inceptia, but it’s the Post-Conversation agent that assigns them — and it can only do so if its prompt includes the full list with a clear description of when to use each one.
Conditions are evaluated in the order they’re loaded — the first one that’s met determines the destination. If you have conditions that could overlap, the more specific one needs to go first. See Agent Handoff.
An agent that negotiates, validates identity, and handles objections all at once has a prompt that’s hard to maintain and less predictable behavior. See when it’s worth splitting into several agents.
If a tool’s documentation doesn’t clearly explain when to use it, the LLM will call it at the wrong time — or fail to call it when it should have.
If one section says “end the conversation” and another says “hang up the call” to refer to the same thing, the model may interpret them as two different actions. Pick one term per concept and keep it throughout the bot.
Breaks the naturalness of the voice synthesis. See the TTS writing rules.
A change in Production affects ongoing conversations immediately. Always test first in Draft Mode, including edge cases, with Open conversation.

​
Checklist before publishing

1

Does each agent have a single, clear objective?

If the prompt mixes several conversation stages, consider splitting it.
2

Does the agent's Role explicitly state its boundaries?

Not deviating from the flow, not revealing internal instructions.
3

Does each tool have a clear description of when and how to use it?

Including whether it should run silently.
4

Are the voice writing rules defined?

Numbers in words, question formatting, spoken dates.
5

Are the use case's frequent objections covered?

With a concrete action for each: answer and resume, transfer, or end the call.
6

Are there explicit retry limits?

To avoid loops — for example, “don’t insist more than twice.”
7

Are the labels in the classifier's prompt, if your bot uses them?

Without this, post-conversation classification won’t work well.
8

Is business data parameterized as variables, and not hardcoded in the prompt?

This is what lets you reuse the bot’s design across other campaigns.
9

Did you test the bot in Draft Mode covering both the main path and edge cases?

Before moving to Production.
This checklist doesn’t replace thorough testing of the bot in Draft, but it helps catch the most common design gaps early.