I created a chatbot for my online store but no matter what I do, it’s buggy.
I’m using Claude Sonnet 4.
The prompt will not follow the basic chatflow instructions. For example, I first instruction is lead capture, get Name and email before continuing the conversation. It follows this 50% of the time. In the anthropic playground, it works well EVERYTIME.
What using open ai (I can’t even use openai), it displays what should be back information to the user. Like part of the prompt, it’s thinking, refers to the user in the 3rd person.
I’m tired of testing it, the problem is not the ai integrations, it’s the platform.
I increased the memory after reading another post that someone had similar problems, it didn’t help.
On top of that, the chatbot keeps repeating itself. Asking questions it already asked. Refusing to end the conversation when it’s clearly prompted to end the conversation. I’m tired.
From what I’ve read, Claude 4 models are more literal and benefit from very explicit, step-by-step instructions. Try making your lead capture flow even more explicit, for example:
“First, ask for the user’s name. Wait for their response.
Next, ask for their email address. Wait for their response.
Only after both are received, continue the conversation.”
Also, it can help to provide examples of the desired flow in your prompt. Sometimes, showing rather than telling helps the model understand what you want. I hope this helps!
Hey @roxie best practice is to write your System prompt using Markdown format.
I use the following approach for all my clients and personal project, and it never fails:
##Role
You are an awesome assistant...
##Rules
1) Always refer to the following KB files when answering...
2) Always use bullet points when summarizing...
When you say you increased the memory, I assume you’re referring to the ‘Memory buffer’, correct? If so, that will not affect your issue at all, the MB is for allowing the chatbot to remember chat history from {{how many tokens ago}}.
Ensure you are not using the following sections (you don’t have any instructions in them) as this may cause conflict in how your AI functions:
Model reminder
Training dialogue
Regarding redundancy, you want to make sure you’re not repeating instructions or adding any conflicting instructions within the prompt. The more clear and concise your system prompt it, the more effective your results will turn out.
Regarding ending the conversation at a desired stage, this has to do with how you programmed your conversational logic. Make sure to add a clear and sequential conversational logic under the ##Rules section of your prompt for your AI to mimic.
If these don’t work, and this project is very important to you, send me a DM with screenshots so I can see what else I can possibly do.
The tool is functioning as expected. From my testing (screenshots attached), the chatbot won’t proceed without collecting the email and insists on getting that information, which, as I understand, is precisely what you intended.
If this trend continues, it appears that your chatbot may benefit from a few prompt adjustments. I noticed you want it to use specific terminology, which you’ve mentioned in the prompt section. You might want to move that instruction (for example, “you MUST say exactly…”) into the Model Reminder section, which is located just below the Prompt section. Model reminders are shown to the model before every single reply, so they help keep responses consistent. I would recommend clicking on “Introductions” as well and keeping the desired introductions there.
@Ned.Malki already made some great suggestions worth noting:
Avoid redundancy: make sure you’re not repeating or adding conflicting instructions within the prompt. The clearer and more concise your system prompt is, the better your results will be.
Control conversation flow: if you want the conversation to end at a certain stage, add a clear and sequential conversational logic under the #Rules section of your prompt so the AI can mimic it.
I also noticed that you have both a knowledge base and a prompt, such as “MUST NEVER recommend products not found in your database search.” That’s fine, but to make it work as intended, you’ll need to adjust your Knowledge Base settings (specifically, lower the Relevance Cutoff so responses stay within your knowledge base).
I am attaching screenshots when I tried to use your tool and it looks like it doesnt deviate from conversation without colelcting email and insists on getting that information (my understandhing is thats exactly what you want)
If you encounter any errors or issues, please email info@pickaxeproject.com so we can do deeper troubleshooting. Thanks for being part of the Pickaxe community!
