When we set up our AI assistant (built with the AI Chatbot widget), we used ChatGPT to help structure and refine the instruction wording — but most of it still had to be written and tweaked by hand.
A lot has changed in the AI Chatbot setup experience since then, so I’d really love to hear how people are approaching this today.
Are you using LLMs to help train your chatbot, or do manually written instructions still work better for your use case?
Would love to hear what’s been working for you so far!
Yeah, we did use ChatGPT (usually the latest version) to refine the wording and make the instructions clearer for the assistant. That helped us reduce ambiguity and keep the overall structure and logic more consistent.
Another interesting use case was asking ChatGPT to review the instructions themselves, mainly to spot things that could be improved, shortened, or clarified.
Would really love to hear how others are handling this and what’s been working best for you so far
I prefer to do it manually. We found that when you keep the instructions more “human” the AI is more creative and gives more quality, in context responses. Sometimes it is tricky and the AI goes off topic, but hey, does a human not do the same ;-). If you want to go away from “perfect responses” and have a creative AI for us only manual storytelling like introductions work best.
@Ralf_Gerhardt, great insight about “human” instructions leading to more natural responses!
Keeping the structure with a storytelling approach must be tricky as well, as you’ve got to moderate your own creativity to not go off on a tangent yourself
Have you figured out a go-to template for manual instructions, or do you maybe use an LLM just for the structure, while keeping the content itself manual?
Good question. I really try to treat the AI like a human, went away from the idea of temolates, like I don’t have a template speaking to my wife or kids - if that makes sense ;-).. What I do is keeping the short term memory (last 15 chats) and the knowledge base on topic and current all times. Actually using old school project management to keep my AI in context. For example my knowledge base consists of targets and once they are done I mark them done and the file gets archived, which means that AI does not access it anymore. Not sure if this makes sense. It is actually quite simple. You simply remove outdated content off the knowledge base and add new one once needed for the project. It is a different approach going away from working with a machine rather treat the AI like a strategic and tactical partner who always is in context about the project or task you try to solve There is more to it, but maybe this helps
I wrote four one page documents each covering a different area.
I wrote a short summary doc - some instructions on what was my Agent’s purpose, and that perusal of the docs each time was essential.
I explained that the docs have version numbers, and can be updated at any time.
I also very specifically stated what I wanted said if my pal didn’t have a clue about what the user was talking about - basically saying they should scroll down to the bottom of the page and complete a request on the form there, and a human would definitely assist. And that’s how I’m genuinely able to have my Agent proclaim at the commencement:
“What’s up? I guarantee 100% that I can help you…”
The type of AI chatbot training required depends on the complexity of the website, what it offers (i.e., the type of business), and its built-in features. In my case, I trained Elfsight’s AI chatbot using both manual and automated (i.e., LLM-based) methods. In the end, the instructions developed with ChatGPT worked best.
Why? Because Elfsight’s AI Chatbot runs on GPT-5 mini, which appears to better understand the command language developed by its parent company. Given the complexity of our website and what we wanted to achieve, it took more than 800 interactions to get everything working correctly.
Now, users can visit our site and ask the chatbot virtually anything related to aviation and aerospace, and it responds appropriately. Additionally, if someone submits questions related to security violations, copyright infringement, or employment matters, the chatbot responds properly and in alignment with our intended guidelines.
Bottom line: I highly recommend experimenting with both manual instructions and LLMs (e.g., ChatGPT, Gemini, Copilot, and Grok) when developing your AI chatbot. Also, stay focused on creating training instructions that align with your business and the specific, detailed goals you want to achieve.
Useful Tip: To speed up your chatbot training efforts, ask the LLM to return its instructions in markdown-ready copy-and-paste format. This makes it much easier to safely transfer instructions containing HTML, code snippets, and other structured content directly into your chatbot’s instruction fields.
@Ralf_Gerhardt loved your “no templates” approach, Ralf - makes total sense!
I also like the way you’re handling the knowledge base dynamically instead of trying to keep everything permanently accessible. Feels much closer to how humans actually work with context too
@user16632 cool to see that manual training seems to work best for many users here!
Providing clear instructions and covering situations where the chatbot doesn’t have enough context definitely seems worth the extra effort. And a concrete fallback guide for moments when the chatbot gets “lost” is a really smart move as well
Waiting for your use case insights once everything goes fully live!
@Petar_Dietrich, your comments are super insightful as always, thank you for such a detailed breakdown!
Your point about website complexity is absolutely fair, and I really like your recommendation to experiment with both manual and AI-assisted approaches to see what works best for a specific use case.
Thanks a lot for sharing this, Petar! It’s fantastic to hear such hands-on insights from real user experience
I’m just learning vibe coding and will probably adapt some new workflows to my AI chatbot, too. First, I uploaded all documentation to a Google NotebookLM notebook. This allows the AI to sift through everything I have and draft relevant, knowledge base prompts for the AI chatbot. It should allow for analysis of what’s appropriate and what needs refinement so that AI can understand the directions. I’m sorry, but there’s a big difference between human understanding and AI understanding. I talk to family because they’re humans. AI has definitely not reached that level. It’s more like an Autistic human. It’s brilliant at some things, but if you get anywhere out of its comfort zone, it can freak out until you give it what it needs.
Also, I just read about Google Lab’s Opal today. So, I’m going to try that by using Opal to create a mini app specifically created to improve or create prompts. It’ll be interesting.
Feeding all documentation into NotebookLM for prompt creation sounds like a super smart workflow. Was that what inspired you to think about building your own app around it?
Would really love to hear how that experiment goes for you
