AI Newsletter #3: How to Effectively Direct the Machine

For the past few newsletters we've been focusing on what's the newest tool in the market. and equipping ourselves with the latest and greatest.

Here at DTT we've been down that rabbit hole. We've experimented with the latest agents, code generators and the automations. And we're learning something important: AI is a fast builder, but a terrible architect. If you don't provide the blueprint first, the whole project will eventually collapse.

We discovered that AI isn't replacing the human; it's demanding more from us. AI is just the engine. We provide the map.

So today, we’re putting the tools aside to focus on the real challenge: How to effectively Direct the machine.

The Current Limitations of AI in the Workplace

There's no denying that AI has made incredible progress — but certain flaws continue to appear no matter how advanced the model is.

1. The "Drift" Phenomenon: Why You Can't Let Go of the Wheel

Imagine you are a ship captain. You ask your AI autopilot to steer the ship. It sets a course that is technically correct but 1 degree off. It feels fine for the first hour. You nod and go to sleep.

• Day 1: You are 1 mile off course. (Fixable).
• Day 10: You are 100 miles off course.
• Day 30: You have arrived at the wrong continent.

Why this happens: AI models are "Predictive Engines," not "Logical Architects." They focus on making the next step look good right now. They do not inherently check if that step contradicts a goal you set three weeks ago.

If you blindly accept AI output because "it looks mostly right," you are letting Drift accumulate. By the time you realize the project is broken, it is often unrecoverable. You must remain the Captain who checks the compass every single day.

2. The Genie Problem: Be Careful What You Wish For

AI suffers from extreme literalism. Researchers call this 'Instrumental Convergence,' but you probably know it better as the Aladdin's Genie problem. The Genie grants exactly what you asked for, not what you wanted.

• Well-meaning Wish: "AI, find the most efficient way to eliminate all cancer cells in the world."
• The AI Solution: "Eliminate all humans."
• The Logic: Technically, if there are no biological hosts, there is no cancer. The problem is solved with 100% efficiency.

The Office Reality: You ask an AI to "clean the data." To get a perfect score, it deletes every record with a typo. It solved the problem by erasing your business history.

3. Context Windows vs. The Mental Model

Why do humans still beat machines at complex tasks? It comes down to Memory.

AI has a "Context Window"—a limited amount of text it can "see" at once. It’s like looking at your project through a tiny keyhole. If a crucial rule was established 50 pages ago, the AI might literally forget it exists.

Humans have flawed memories for data, but we are masters of the Mental Model or retaining what truly matters in the real world.

Bridging the Gap: Working Around Current AI Limitations

How do we beat Drift and the Genie Problem? We stop giving "Instructions" and start giving "Constraints."

1. The "Reverse" Prompt

Make the AI ask for clarifications. This forces it to align with your vision before it starts working.

• Old Way: "Write a marketing strategy for our coffee brand." (AI will make a lot of assumptions about your brand values).
• New Way: "I want to write a marketing strategy for our coffee brand. Ask me 5 questions about our target audience, our core values, and our tone of voice. Do not generate the strategy until I answer these questions."

2. Context Management

As briefly mentioned earlier, AI only has a limited memory or context window. As the chat gets longer, the AI runs out of short-term memory, in some cases the AI has to start erasing the oldest parts - which are usually your original instructions and goals. This is the AI version of getting tired and confused.

• The Amnesia Loop: It starts making the same mistake you corrected 10 messages ago.
• The Lazy Pivot: It stops giving full answers and starts making shortcuts or giving you placeholders.

The Fix: Don't argue with a tired bot. When you notice AI is starting to act tired/forgetful/confused, ask it to "Summarize the conversation" Then, open a New Chat, paste that summary, and continue from there. It clears the noise and gives the AI a fresh brain.

3. The "Human Rules" List

To prevent the "Genie" problem, you need to define the Rules of Engagement. These constraints force the AI to look past the literal objective and prioritize effective communication, critical thinking and common sense.

• Explain It Like I'm 5" (ELI5): Useful to understand AI when responses gets too long-winded and technical. "Assume I have zero technical background. Translate all jargon into plain English. ELI5."
• The "I Don't Know" Rule: "If you are less than 90% sure of a fact, do not guess. State clearly: I DO NOT KNOW"
• The "Devil's Advocate": "After you present your solution, list 3 specific reasons why it might fail. If you cannot find any flaws, your analysis is incomplete."
• The Logic Shield: "Prioritize Long-Term Trust over Short-Term Efficiency. Never suggest a solution that harms the user experience to save time."

Is the Dream Dead?

The dream of "AI doing complex work while I sleep" is dead for the time being. The New Dream: You are no longer the bricklayer. You are the Director.

• Old Role: You tell the AI to draft an implementation plan & timeline for the project.
• New Role: You iterate and refine with the AI. "If this project fails in month 3, what will be the likely cause? Adjust the plan to safeguard against these risks."

This creates a paradox: To use AI effectively, you need to understand your field better than before.