Last week, working at home by myself, I did something that was clearly irrational—I yelled at ChatGPT. We’re talking a full-on, completely exasperated, “OH FOR <insert your favorite phrase here>!!!”. I had to step away for a moment to calm down. Later that day, the real lesson finally landed.
I was trying out new software and couldn’t figure out how to enable a specific feature. When I asked ChatGPT, the steps it returned were a dead end—the menu it suggested was nowhere to be found. Next, I told it to search the internet for a different answer. That path also led nowhere.
My to-do list was long. I was testing the tool because I thought it would save time. So you can imagine my frustration building when my AI “co-worker” seemed asleep on the job. I tried one more time and told it to read the software company’s documentation. I switched it to Thinking mode. About a minute later, it gave me… the same answer as before. Hence, my momentary mini-breakdown.
Later, after resetting and working on something else, I remembered the software had just released a major new version. ChatGPT likely didn’t know I was using the latest release. Sure enough, once I provided that context, it gave the right answer and I was on my way. I’m sure everyone has felt that sheepish moment when the problem was missing input, not bad performance.
That’s the point: AI is now so capable that wrong answers are often a context problem. Knowing that is easy. Acting on it is harder. At Transparent Partners, we’ve been focusing on providing context to LLMs in several practical ways. So it’s a good time to ask: Is your enterprise AI strategy maximizing these five context layers? If not, let’s discuss how we can help you build a more effective roadmap.
1) The Immediate Context: Role, Goal, and Constraints
The most accessible layer of context is prompt engineering. Many people still treat LLMs like search engines, entering a few keywords and hoping for the best. A better approach is to prompt the way you’d delegate to a co-worker.
Use the “Role, Goal, and Constraints” framework. Instead of “Write an email about the project update,” try: “Act as a Senior Project Manager (Role). Write a project update email to stakeholders summarizing delays in Q3 caused by supply chain issues (Goal). Keep the tone professional but reassuring, and limit it to three bullet points (Constraints).” This narrows the model’s search space to the slice of reality you actually mean. In other words: the first of the enterprise AI context layers is simply delegating clearly
2) The Data Context: Leveraging Large Context Windows
Sometimes description isn’t enough; the AI needs to see the source material. Modern LLMs can hold a lot of information in their working memory. That lets you move beyond short prompts.
Rather than summarizing a 50-page vendor contract or typing rows of sales data, upload the PDF or spreadsheet into the chat. Then the model can answer from those files, not from general training patterns. In practice, this turns a generalist into a specialist on your documents.
3) The Persistent Context: Custom Instructions, GPT’s, and “Gems”
If you keep repeating the same preferences—“I work in HR,” “Keep it concise,” “Format as a table”—you’re wasting cycles. Persistent context solves that. It’s available through ChatGPT custom instructions and custom GPTs, or Gemini’s “Gems.”
These tools let you bake your defaults into the background of every interaction. You can create an “Executive Assistant” that follows your formatting rules. Or you can build a “Brand Editor” that understands your voice is witty but professional. That way, each new chat starts with the basics already loaded.
4) The Retrieval Context: RAG (Retrieval-Augmented Generation)
For enterprise use cases—or for data that changes daily—manual file uploads don’t scale. This is where Retrieval-Augmented Generation (RAG) comes in. Think of RAG as an open-book exam.
Instead of relying on the model’s internal memory, the system searches your live sources first (a database, intranet, or knowledge base). Then it feeds the most relevant facts to the model alongside your question. This anchors output in your current, proprietary data and reduces errors—especially when policies and numbers shift often. If your use case fits this pattern, getting help to implement RAG (or reaching out to us) is usually worth it.
5) The Agentic Context: Multi-Step Reasoning
The final frontier is agentic workflows. Here, the AI doesn’t just answer; it decides what to do next. With the context “Schedule a budget review with the finance team for next Tuesday,” an agent can infer a sequence: check the current date, scan calendars, book a room, and send invites.
In this setup, context becomes the logic engine. It tells the system what tools to use and what steps to take. That’s how you move from a passive chatbot to an active digital employee and why the fifth of these enterprise AI context layers often changes operating models, not just workflows.
I hope these suggestions make it easier to provide more context in your AI interactions. And if you still get an incorrect result, remember: yelling at an inanimate object won’t get you there any faster!
Dan Weinbeck, Director
