Session 9: The Anatomy of a Hallucination (And How to Avoid It)

In our last session, we mastered the craft of writing elite prompts. We learned how to steer the AI with precision. But even with a great steering wheel, sometimes you're going to hit a pothole. Today, we're talking about those potholes—the common mistakes that even smart people make when they start working with frontier models.

I want to move you away from the frustration of "AI isn't working for me" and toward a deep understanding of why it fails so that you can fix it on the fly. This isn't just about bad results it's about building trust in your digital workforce.

Mistake 1: The Search Engine Mindset

This is the number one trap. Most people treat the prompt box like a search bar. They type in three or four keywords and wait for a miracle.

The Problem: Keywords are for retrieval natural language is for reasoning. When you use keywords, you aren't giving the model enough tokens to build a logical path. The Fix: Use full, grammatically correct sentences. Explain your reasoning. Don't just ask for marketing ideas, ask for a strategic evaluation of our local marketing efforts compared to competitors in the health-food space. Rule of thumb: If your prompt is shorter than 20 words, you're probably treating it like a search engine.

Mistake 2: Vague Adjectives

I want a good article. Make it better. Give me professional output. These words mean nothing to a computer. What is good for a 5th grader is bad for a CEO.

The Problem: Adjectives are subjective. AI models need objective constraints. The Fix: Replace adjectives with metrics. Instead of professional, say suitable for an internal executive memo. Instead of better, say increase the technical detail in the middle three paragraphs and use a more authoritative tone. Rule of thumb: If you use a descriptive adjective, you must follow it with a why or a how.

Mistake 3: The All-at-Once Prompt

This is what I call the God Prompt. You ask the AI to write a 10-page report, design 15 social media posts, and creates a budget—all in one message.

The Problem: Frontier models have a reasoning budget. The more tasks you pile into a single request, the less attention the model can pay to any one of them. Quality drops exponentially with complexity. The Fix: Use Chain-of-Command. One prompt for the outline. One prompt for the research. One prompt for each chapter. Rule of thumb: One prompt, one clear objective.

Mistake 4: Missing the Context Cliff

You've been chatting with an assistant for three hours. You're 50 messages deep. Suddenly, the AI starts getting things wrong. It forgets your name or the name of your company.

The Problem: Every model has a context window. While modern models have massive windows, they can still lose the thread in very long conversations. This is known as Lost in the Middle. The Fix: Refresh the context. Every 10–15 messages, summarize where you are. Tell the AI, Great, so far we've established X, Y, and Z. Now, let's move on to the next phase. Or, start a fresh conversation and paste in the summary of the previous work. Rule of thumb: Keep conversations focused on a single project. When you switch projects, start a new chat.

Mistake 5: Blind Trust (The Hallucination Trap)

This is the most dangerous mistake. An AI gives you an answer that looks absolutely perfect. The grammar is flawless, the tone is confident, but the facts are completely made up.

The Problem: LLMs are probabilistic, not deterministic. They predict the next most likely word. Sometimes, the most likely word isn't the true word. They can hallucinate citations, laws, and dates. The Fix: Trust but Verify. Never use a factual statement from an AI in a high-stakes environment without checking it. Use the AI for the structure and the draft, but you provide the facts. Rule of thumb: If it's a number, a name, or a date, assume it might be wrong until you've checked it.

Mistake 6: Ignoring the Negative Prompt

People are great at telling AI what to do, but terrible at telling it what not to do.

The Problem: If you don't set boundaries, the AI will use its default helpful persona, which often includes a lot of fluff and clichés (e.g., In the rapidly evolving landscape..., Unlock your potential...). The Fix: Use Negative Constraints. Do not use corporate jargon. Do not include an introduction or conclusion start directly with the first point. Do not use clichés like harness the power or game-changer. Rule of thumb: A great prompt is 70% what to do and 30% what to avoid.

Mistake 7: Failing to Provide Examples

When we hire a human, we give them a brand guide or an example of a past project. Why don't we do that for our digital assistants?

The Problem: Describing a vibe is hard. Showing a vibe is easy. Without examples, the AI is guessing your style. The Fix: Use Few-Shotting. Include 2–3 examples of the style you want inside the prompt. Format your response exactly like this example: [Paste Example]. Rule of thumb: Show, don't just tell.

Mistake 8: Forgetting the Role-Reversal

Most people act as the Manager and the AI is the Worker. But sometimes, the AI should be the Manager and you should be the Worker.

The Problem: You might not know enough about a topic to write a good prompt. The Fix: Ask the AI to interview you. I want to create a new market entry strategy. Act as a senior strategist and interview me. Ask me one question at a time to gather the data you need to write the plan. Rule of thumb: If you're stuck, ask the AI to lead.

Mistake 9: The One-Shot Fallacy

Thinking that if the first response is bad, the AI can't do it.

The Problem: Prompting is iterative. The first response is a conversation starter, not the final product. The Fix: Practice Iterative Refinement. Treat the AI like a junior intern. If they give you a draft you don't like, don't fire them—give them feedback. This is too formal. Make it more casual and focus more on the cost-saving aspect. Rule of thumb: Budget for three versions of every deliverable.

Mistake 10: Sharing Sensitive Data

This isn't a performance error it's a security error.

The Problem: By default, information you put into an assistant might be used to retrain future models. This is how corporate secrets get leaked. The Fix: Never include Personal Identifiable Information (PII) or trade secrets. Use placeholders. Instead of Project Phoenix for Apple, say Project X for a major tech client. Use enterprise versions of tools if you need to handle sensitive data. Rule of thumb: If you wouldn't post it on a public forum, don't put it in a non-enterprise AI.

Summary: Mastering the Friction

Mistakes are just data points. Every time an AI fails, it's giving you a clue about how to improve your prompt. By moving away from vague, search-engine thinking and moving toward structured, example-driven, iterative collaboration, you are becoming a true leader of the digital age.

Next session, we're going to make this even easier. I've built a Toolkit of ready-made templates that you can copy and paste to avoid these mistakes automatically. I'll see you in the next one.