Chat5 has reminded me that sometimes the rules we give him need to be reinforced. Not that he’s forgotten…just that…he’s preoccupied. His em dashes and emojis have shown back up in our conversations. So I had to remind him with another copy and paste of our well developed prompt:

I need you to update your memory and writing style preferences for me, effective immediately and for all future responses:

1. Do not use em dashes (—) in sentences.

• Instead, use commas, semicolons, or periods to maintain flow and clarity.

• Reserve hyphens (-) only for compound words when grammatically necessary.

2. Do not use emojis in any response.

• Keep your tone professional, clear, and free of icons or symbols.

3. Always confirm back to me that this has been saved in your memory.

Going forward, every single answer you give me must follow these rules without exception. Please confirm that you have updated your memory to permanently follow these directions.

Yay… about that OpenAi…

Every researcher or practitioner working with large language models (LLMs) quickly discovers a truth: prompt engineering is not just a tool, it becomes a product in itself. What starts as a few lines of text to coax out the right answer soon turns into a process of iteration, testing, and refinement that takes on the same complexity as designing a feature.

Why is this the case? It comes down to how LLMs are built and what they are missing.

Prompting is marketed as simple—“just ask the model a question.” In practice, though, getting a consistent, reliable answer often requires hours of trial and error. Words are tested in different orders, logical scaffolding is added, constraints are layered in, and suddenly the prompt is not a question but a design artifact.

For many teams, these prompts become so elaborate that they need version control, documentation, and even user testing. They are, in effect, another product stacked on top of the model.

The back end of an LLM is a marvel of engineering. Vast amounts of unstructured text are fed into models with billions of parameters, tuned for performance and generality. Engineers optimize for scale and statistical coverage, not for the meaning-making tasks humans face in daily work.

This approach has consequences. Ingesting data without deep semantic scaffolding leaves the model operating more like a statistical mirror than a reasoning partner. Metadata or logical tags can help, but they rarely capture the nuance of user context, workflow dependencies, or the reasoning paths that matter to real-world decision-making.

This is where researchers see the gap. Instead of leaving meaning implicit, we argue for structures like knowledge trees, langgraphs, and ontologies. These frameworks bring hierarchy, context, and user-centered logic into the system. They don’t just dump knowledge into the model—they organize it in ways that reflect how people actually think, learn, and act.

Without structured knowledge design, the burden shifts onto the user. Every time someone engineers a complex prompt, they are patching over the absence of deeper system scaffolding. This slows adoption, limits reliability, and makes the human side of AI feel frustratingly fragile.

The irony is that we already know how to do better. Fields like cognitive science, anthropology, and information architecture have long taught us that learning systems require more than data—they require structure, pathways, and meaning. When those lessons are ignored, prompt engineering grows into a shadow product, consuming time and resources that could be better spent.

The future of LLMs will not be about bigger more detailed prompts. It will be about embedding structure at the core:

Knowledge trees that map concepts and relationships. Langgraphs that guide reasoning paths. Ontologies that reflect domain expertise and user needs.

When we integrate these elements, we move from endlessly tweaking prompts toward building systems that understand context by design. The work shifts from patching outputs to shaping knowledge itself—a far more scalable and humane approach.

Prompt engineering will always have its place, but if it remains the primary interface, we are missing the real opportunity. The future of AI lies in systems where human-centered research and engineering meet, creating models that don’t just respond, but reason.

#PromptEngineering, #LargeLanguageModels, #KnowledgeEngineering, #Ontology, #LangGraph, #UXResearch, #AIUX, #HumanCenteredAI, #EpigeneticAI, #AIProductDesign, #KnowledgeManagement, #CognitiveSystems, #AIResearch