Problem

When I was tasked with design a chatbot (prompt), I wanted mine to be able to execute certain commands or tasks within the conversation based on specific trigger words. I found that traditional chatbots lacked the ability to perform complex actions, which made it difficult for users to get the help they needed - or get the same learning quality they could have experienced in other settings. Additionally, I saw a lot of chatbots "forgetting"/missing important parts of the prompts.

📄 Learn more about this approach here.

Solution

To solve this, I implemented different (markup language) designs to underline the structure and priorities within my prompt:

I used a clear numbered list of tasks (1., 2., …), ranking the importance of what I want the GPT model to do and how to behave:
1. 1. First, ask the user about ...
2. 2. Based on the reply, make sure that ...
3. 3. Over the conversation, ensure ...
With stars (**) next to words or phrases I highlighted for the AI, what I care most about in the prompt. ChatGPT understands these Markdown markup language codes and is able to prioritize parts of the prompt accordingly:
- Make sure that the **conversation is polite and respectful**.

Trigger words

To solve the issue of complex commands, I got the idea of a “trigger word” to make the AI execute pre-programmed tasks in the conversation: E.g.: If the user types “##Finish##” in the chat, they make it generate a final resume of the chat/ learning session. Thereby creating a personalized learning-summary and handout of the individual learnings and context within seconds. This is especially helpful for complex commands like the description of a specific template or format of the summary, the consideration of defined learning objectives, etc.

Instead of writing detailed prompts repetitively, users can simply use the trigger words to perform pre-defined tasks

This is an example of how I used trigger words and markup language:

**Step 7**: When the user types "##FINISH##", write a summary of all learnings discussed prior in the chat, so they have a personalized learning-card with the most important learned information about Football.

Impact

Significantly improved the usability of my chat bots
Optimized the pedagogical learning experience of my chat bots by highlighting important nodes in my prompt
Giving users the option to trigger certain behaviours based on their personal interaction with the chatbot: Create learning summaries, creating personalized context based options, ...

"This solution has completely transformed the way me and my users interact with our chatbot. It's now much easier for us to get perfectly tailored responses, and we've seen a significant reduction in additional prompts or finetuning withing the chat conversation."

Especially for my (Football) AI chatbot, this solves a big time constrain issue: The users watch a live game and need perfectly tailored responses in real time. By integrating more and more trigger words into the prompt, learners receive increasingly tailored responses and help.

🏈 🤖 Learn more about my Football AI here

Advanced Chatbot Prompt Engineering (Trigger words and highlighter) [Bonus Work]