1. How AI Characters Think and Speak
To engage in conversations with learners, the AI Characters (or AICs) you build with Wonda use a Large Language Model (LLM) to generate responses.
An LLM is a predictive AI model trained to follow instructions and generate natural language replies.
By default, Wonda Pro uses the gpt-4o-mini from OpenAI.
Here’s what happens behind the scenes for each learner message:
For each message that an AIC receives, its response will be generated by the LLM using a) the AIC Prompt you have defined and b) the conversation history between the learner and the AIC gathered during a given session;
Once the response has been generated, it is displayed instantly in the chat window (as a text message) and transformed into an oral response using the Voice synthesis model.
💡 If you’d like to use the full gpt-4o model, switch to a different one, or integrate your own LLM, please contact the support team.
2. Best Practices and Tips
When designing your simulation, you’ll prompt the AIC in the Simulation Editor.
Prompting is similar in many ways to writing a detailed brief for a comedian where you can decide to make it super directive or, to the contrary, leave a lot of room for improvisation.
While prompting is not an exact science, here are a couple tips that might help you get better results instantly.
Break your prompt of your AI Character, into specific categories.
We recommend in most our templates the following categories:
Who you are
Who you are interacting with
What you know
What you role is
Important things you should do
Do not hesitate to add surprising details about your AIC backstory or personality traits.
By going into details in the “Who you are”, you will be able to turn the default LLM behavior (a helpful assistant for basically anything), into a character that“feels” more human and most importantly, is uniquely relevant for the conversation you want to simulate.
Avoid long-winded or overly vague instructions and use numbered list for discussion steps.
Keep in mind that LLMs can be quiet good at following instructions provided those instructions are structured and directive.
Start with a small prompt, speak to your character and then edit the prompt.
The best moment is when you start seeing your AI Character come to life in front of you while speaking to him and iterating with its prompt. As a rule of thumb, it can take about 10 iterative cycles (about 1h of discussion in total) to start getting the results you would like.
We also recommend a prompt of 15,000 words or less. Keep in mind the shorter the prompt the more each instruction will be followed correctly
Interested about how to analyse future discussions with your AI Character? Here is How to Write Good Criteria to Evaluate Conversations and provide feedback to your learners based on their performance.