Based on the experience I have gained through testing, trial and error over the past few months, I have discovered that there are two types of Skills in Claude—or two types of GPTs in ChatGPT—that interfere with or undermine each other. What I was trying to do was build Skills or GPTs that would first explore certain topics with me and then generate documents, artifacts, analyses, or research. I now understand that the two tasks I was asking them to perform were antagonistic, and that their instructions were contaminating one another within the AI’s context. I have given them names to make the distinction clearer: - Dialogue Skills — discovery, facilitation, exploration, Socratic dialogue. - Production Skills — document creation, content tools, research, and analysis. The first type should not make assumptions, search the internet, or propose actions. It should only listen, summarize, and clarify. I see this as useful when I want to clarify what I am doing or thinking. In such cases, it is counterproductive for the AI to search online, make assumptions, or ask me to choose from options that it has generated itself. This type is also useful when I am developing my brand voice. The second type needs to search online for figures, statistics, research, and factual data in order to provide solutions. It is useful, for example, when I want to calculate TAM or SAM. What I discovered is that when I include instructions from both categories in the same Skill or GPT, they interfere with one another. I end up with neither the best dialogue and exploration nor the best research and analysis, precisely because the model is operating with contradictory instructions in its context. I tested splitting the process into two separate Skills or GPTs that run sequentially. The first, focused on dialogue, produces a document that then serves as the source for the second. The results were significantly better: a genuinely Socratic dialogue, unaffected by unsolicited proposals and recommendations—which are often tempting—followed by a much more factual analysis that required far fewer input requests from me.