In the first episode, I mentioned I started this podcast while on vacation. During a particularly brutal airport delay, I expanded my research into clear instructions. The AI had an amusing cake order scenario to illustrate how to add more context and richer instructions. This reminded me of an example from my daughter’s childhood I wanted to  share in this second podcast. Also please note, in case you want to download the materials and full transcript, I'll add a link in the Show Notes.

✅ How to prevent AI hallucinations in complex data tasks

Providing clear instructions to AI involves two things: how you describe the task itself, and how you format and structure the instructions. When describing the task, my advice is simple: more is always better.

✅ How to use an LLM to improve and optimize AI instructions

The beauty of AI is you can ask the LLM how to improve your instructions—a great way to learn how AI thinks. I tend to use different instructions for different tasks. For general tasks I do not go into much detail, but as you will see in the later chapters on Python, HTML, and drawing images, I am very detailed. Consider using these categories when crafting your instructions:

- Your persona: Setting a clear role or perspective for the model to adopt.

- Background or context: Giving the baseline environment or history for the task.

- The specific task: Stating the exact action you need executed.

- Negative constraints: Explicit instructions on what the AI should strictly exclude.

- Structural boundaries: Defining specific delivery rules, such as organizing text into sections without altering the source wording.

- Validation logic: Requiring visual markers—like a green checkmark before a suggested change—and a summary of proposed changes.

- Before-and-after data examples: Clear samples of raw data input paired with your exact expected output. For these, I use Markdown tables or CSV files.

✅ What is context window friction and how do tokens affect LLM focus

AI also suffers from 'context window friction'—which is one way to say it runs out of tokens and tunes you out. When an interaction gets too long and complex, AI sometimes hits a boundary where it struggles to anchor its immediate focus. At that point, AI turns into the teenager staring at their phone and ignoring you. To keep system instructions but clear the erasable chat history, there are two instructions I use in this situation.

-               If you run low on tokens, stop and warn me.

-               Forget everything before this moment. Moving forward, focus strictly on [New Task] using [Specific Format].

-              If it looks wrong, ask the AI to show you the original source of an answer. Often it will say it can't because it made it up.

✅ Why does AI prefer structured markdown over unstructured text

In this AI era, I think the most powerful thing you can do is to shift your writing into pure, structured plain text, because AIs—or Large Language Models—run on plain text. When we feed AI tools messy, verbose data they hallucinate, give wrong answers, lose focus, and fail at tasks. When you add structure and formatting to your plain text, you are talking the exact same language as the machine.

Markdown is simply plain text with extra characters to add formatting. For example, adding a hashtag to the beginning of a sentence indicates a level one heading, while wrapping two asterisks around text makes it bold. Markdown formatting includes text that is bold, italic, headings, lists, tables, and emojis.

Consider using Markdown for your interactions with AI. For example, if I want to add an email draft to the conversation, I copy it from Word and paste it into my Obsidian Markdown editor, which automatically converts the basic formatting into Markdown. From there, I copy the structured Markdown text into my AI window. Converting operating manuals or procedures into Markdown in this way is a simple way to ensure the AI uses the data efficiently.

To see the power of Markdown in action, we are going to step through two distinct rounds of talking to AI: one using messy prose, and one using pure logic formatted in Markdown.

✅ Unstructured prose prompt example failure Round 1

In Round 1, Sarah's unstructured email to the bakery's AI ordering system can accidentally trigger hallucinations.

⚠️ The Input

"Hi! Need a high school graduation cake for Chloe. Tennis theme, big racket on top, write 'Congratulations Chloe!'. Wait, her favorite color is neon green, make the tennis balls that bright neon color so they pop. For flavor, she hates vanilla, so definitely make it chocolate with fudge filling. Actually, put 'She did it!' on the cake right below the racket, but make sure 'Congratulations' is bigger. Also, she is allergic to peanuts!! Do not put peanuts anywhere near it, use almonds if you need a crunch. Let me know if that works!"

❌ The Output

The bakery's automated kitchen system prints out a design ticket, and the baking robots create a monstrosity:

-               The Inscription: Written in massive, shaky icing letters across the entire top of the cake is: "SHE DID IT! RIGHT BELOW THE RACKET BUT MAKE SURE CONGRATULATIONS IS BIGGER."

-               The Design: A massive, neon-green fudge-covered tennis racket.

-               The Flavor: A plain vanilla cake (because the AI prioritized the high-signal token "vanilla" and associated the word "hates" as a generic human modifier it didn't know where to apply).

-               The Allergen Disaster: Because the sentence contained both "peanuts" and "almonds" in a flat structure, the AI split the difference and added peanut butter cups around the rim as a decoration.

✅ How to use an IfThenWhy™ approach for structured data logic Round 2

In Round 2, Sarah takes 30 seconds to structure her thoughts using basic Markdown formatting that follows an If Then Why approach.

🛠️ The Input

# Custom Cake Order: Chloe's Graduation

- **Main Theme:** High School Graduation / Tennis
- **Primary Flavor:** Chocolate Cake
- **Filling:** Fudge

## ⚠️ CRITICAL HEALTH ALERT

- **Allergies:** PEANUTS (Strict Zero-Tolerance)
- **Safe Substitute:** Almonds

## Visual Design & Text Layout

- **Top Decoration:** 1x Tennis Racket (with neon-green tennis balls)
- **Primary Text (Large):** "Congratulations Chloe!"
- **Secondary Text (Small, placed below racket):** "She did it!"

✅ The Output

Because of the Markdown structure, the AI engine processes the data with perfect boundaries:

-      The different header levels isolate the text instructions from the data variables.

-      The ⚠️ CRITICAL HEALTH ALERT blockquote flags a high-priority system constraint, ensuring the peanut token is isolated entirely as a negative constraint.

-      The bullet points map the text layout progressively, ensuring the icing decorators write only the specific text strings intended for the cake.

Applying additional categories to the bakery example makes a good thing even better. AI responses improve dramatically with careful guidance like this:

- The Persona: A master baker’s apprentice.

- The background and task: Define the exact action for a custom cake order by combining customer instructions with operational rules like store hours, decorative options, and baking times.

- Negative constraints: Strictly exclude all nuts and cursive script.

- Structural boundaries: Define specific delivery rules and organize complex details—like tiers, flavors, and sizes—without altering the original intent.

- Validation logic: Compare required supplies to current inventory, identify out-of-stock ingredients, and flag them in a final validation summary log.

✅ Why are negative constraints and structural boundaries important in AI instructions

Take, for example, the instructions I gave my two-year-old daughter:

“You cannot draw on any wall, surface, or floor in the house. No crayons, no permanent Sharpies, no markers, no chalk, nothing.”

My daughter demonstrated with brilliant precision where these instructions fell short of good direction. She simply walked out to the exterior covered porch and drew a lovely mural—fortunately in colored chalk and not Sharpies—on every available part of the outside wall she could reach. Since I had not explicitly bounded the rules to include the outside of our home, she assumed that was exactly where I wanted her to create her masterpiece. As a reminder of the importance of better parental instructions, when she went off to college, the mural was still there.

✅ How to use metadata and business rules to guide AI responses

In a future segment, we’ll utilize other plain text formats and languages to organize data even further.  This includes extracting metadata from SQL and leveraging HTML code to retrieve field information. We will also use JSON files to organize business rules alongside physical and logical metadata. These business rules —or semantic data —guide and limit the AI’s response. You don’t need perfectly mapped metadata with data types, normalization, masking or cardinality to get started. Every clue you add increases the odds AI uses the data correctly.

• The Professional Alter-Ego: How to assign the AI a specific Persona.

• The Boundary Lines: Writing clear Guide Instructions and explicit constraints (What *not* to do).

• The Cookbook Method: Providing explicit input and output data examples to prevent hallucinations.

• The Hard Reset: How and when to use "Forget everything we said before this moment."

• Too Much to Handle: Understanding context window friction when the AI tunes you out.