Wondering if you get better results with pickaxe from XML or from other types of formatting? I have noticed that when building directly in LLM apps, such as ChatGPT or Claude, it seems to perform better, and I am wondering how this translates when in Pickaxe.
1 Like
Hi, @aijoe long time no see!
You observation that XML tends to cause your apps/agents to perform better is on point. For standard (non-complex) GPTs, XML provides a machine-readable text that reduces the possibility of hallucinations mainly due to the labeling and paragraphical structure XML system prompts provide. The simpler the agent, the less likely you will need to implement XML or more advanced prompting methods.
A step above XML is POML system prompting (Can be a token vacuum if deployed to the wrong agents and is intended for complex agentic workflows).
System prompting is a science. I break it all down in the Pickaxe Prospectors Association “PPA”
It’s part of the extended Pickaxe community and it’s free to join + free foundational and beginner training.
-Ned