How do you get better and more reliable answers from AI?

To get highly accurate and reliable answers, you must carefully design the layout, whitespace, and visual cues of your request. For example, if you describe an artisanal bakery's allergen list in chaotic, conversational prose, the AI will likely scramble the ingredients—leaving customers to look at the menu and say, 'I'm Not Eating That.' Organizing that information into clean Markdown tables and structural JSON boundaries prevents these dangerous hallucinations and ensures predictable, logic-driven outputs.

How can you improve AI performance and reduce computing costs?

You can improve performance and reduce token costs by using Markdown to keep your data structured and brief. Without this structure, long back-and-forth chats suffer from 'context window friction'—where the AI hits a wall and starts behaving like a teenager who has completely stopped listening to their parents.

How do you make AI generate accurate diagrams, slides, and code?

Instead of writing raw code, act as an editor directing the AI. For instance, you can outline a bakery's daily production timeline in simple, plain words so the AI can instantly generate the Mermaid code to draw a clean, visual workflow diagram for you.

Why does AI lose track of the conversation in long chats?

AI loses track because of a technical limitation called 'context window friction,' where the model gets overwhelmed by too much conversational history. After an extended back-and-forth, the AI hits a wall and starts behaving like a teenager who has completely stopped listening to their parents. To prevent this, you must keep your data structured and brief using clean Markdown, which minimizes context bloat and keeps the AI focused.

The Plain Text of AI Podcast - 1.0 Introduction

Welcome to The Plain Text of AI. I started this podcast while on vacation because I had a lot of free time to research AI and organize my notes. I wanted a podcast because I thought it would be fun to connect and share experiences with you—others who are also wondering about this new AI world. To make your life easier, in case you want to download the materials and full transcript, I'll add a link in the Show Notes.

How to give AI custom instructions

To begin, I’m starting with a practical introduction to the layout, the whitespace, and the structural grammar that makes AI work for you. We’ll also look at the scope, context, and intent of your instructions. Finally, I’ll touch on reference materials —organized in a clean question-and-answer format—that help the model interpret these reference materials.

And keep in mind, these reference materials might be new data provided in your request, or pre-existing knowledge from when the model was trained. I want to establish these foundational concepts now so you have a baseline. In later episodes, I’ll go into the details of each topic. That way, you can choose the topics you want.

How does AI process text

Now, I want to explain why I'm looking into AI at all. My journey into the world of AI started with a foundational question: “How do I prepare data for AI to use?” AI, also called a Large Language Model or LLM, is a model trained to predict and process text. In searching for answers, I discovered that AI lives in the realm of tokens, text files, and structural layout, and is entirely grounded in plain text.

- Tokens are the text chunks AI uses to break down data.

- The more tokens you consume, the more capacity you use. Additional tokens translate to higher costs if you are paying per token via an application interface.

The plain text AI prefers is the raw, unformatted text you type into a basic text editor like Windows Notepad. It is not the text you type in a word processor like Microsoft Word.

Markdown, HTML, and XML are markup languages that use plain text. JSON and CSV are lightweight data-interchange formats written entirely as plain text. Python, JavaScript and C++ are programming languages that are also plain text formats. Mermaid is a JavaScript-based diagramming tool that is a formal text-based language. All of these use plain text that you can edit in Notepad, although each has its own syntax and layout rules.

While the AI responds with Markdown, the chat application web interface itself may render the Markdown into HTML tags. It can be confusing when you paste instructions in Markdown and the AI displays the text without the special Markdown characters. Rest assured that despite the fact AI displays your Markdown text using HTML, under the hood the AI model itself recognizes the structure because it was trained on vast amounts of Markdown-formatted text. It understands the patterns of Markdown formatting natively. We will look at both the input and output choices available in the discussions that follow.

In an upcoming segment, we will also look at Markdown’s structural elements. These include headers, bold or italic emphasis, numbered lists, tables, and precise spacing—all of which help an AI accurately interpret and prioritize information. We have entered a new era of data interaction that requires a different toolset. Instead of using a traditional word processor, the solution is to use a Markdown text editor designed for structure. Markdown files are pure, unformatted plain text that use simple, predictable characters to enforce formatting. Their filenames always end in the extension .M-D.

How to write clear AI instructions

Once you have your text in that clean format, the focus shifts to strategy. How you word your AI requests impacts whether the response is clear and highly focused. You can guide the AI to correct answers using negative constraints and plain-text data examples. We are going to explore how to navigate the physical limits of an AI’s attention span—a concept I call context window friction.

Once you’ve determined what to ask the AI, the next challenge is managing how you lay out the text itself. When you paste rich text from Word into a chat window, Word packages the data into multiple formats: a raw plain text version, a rich HTML version, and often an image to preserve complex layouts. This ambiguity can cause weird spacing, broken lists, and messy text layouts. Worse yet, it floods the AI’s “context window”—its short-term memory—causing the system to get confused and simply stop working.

The soggy pastry

But fixing the physical layout of your text is only half the battle; you also have to give the AI a clear reason to care. Have you ever experienced “Grandma’s Recipe Failure”? You asked Grandma to write down her recipe for your favorite pastry. She listed the ingredients, and the instructions said, “Turn the oven to three hundred and seventy-five degrees and leave the door cracked for 2 minutes.” The pastry came out soggy because you forgot to crack the oven door. Grandma explained the why, “To let the initial steam escape so the pastry edges crisp up beautifully instead of getting soggy.” Grandma understood the baking physics perfectly, but she kept the "Why" as a family secret.

We have all seen procedures that explain if this happens, then do that. But they rarely include the why it matters. There is an old mantra: “If you cannot explain something simply, you do not understand it.” If documentation includes the why, the AI instantly recognizes the passage is highly authoritative, lower risk for hallucination, and fundamentally safe to recommend to a human decision-maker.

It is obvious that the magic lies in Grandma's unwritten know-how. That same type of tribal knowledge is exactly what secures your job in the future.

Shaping your information for AI

An everyday interaction with an AI engine seems straightforward enough—you hand it an instruction, and it hands back a response. Behind the scenes, the AI parses your input against its vast training models. But to get highly accurate, results specific to your business, you may need to provide localized information and raw data assets, such as:

- Databases and structural system schemas.

- Excel spreadsheets and flat data files.

- Medical and technical textbooks containing dense domain knowledge.

- Bibliographies and reference sources.

- Company procedures and internal operational workflows.

While AI can read unstructured prose (standard text paragraphs), modern search and reasoning engines heavily prefer information that is explicitly structured.

Think about a traditional medical textbook. You open a chapter on viruses, find a subheading for 'The Common Cold,' and buried inside a long paragraph is the sentence:

To treat a cold, drink plenty of fluids.

Now, lay out the data exactly how an AI reasoning engine wants to receive it, as explicit, structured key-value pairs. When you map it out, it sounds like this

Illness: The Common Cold.

Treatment equals Drink Plenty of Fluids.

Why: If you ignore symptoms, they could worsen into pneumonia.

In this example, I use bold text to explicitly define the key concepts, and the clean layout removes all conversational fluff. This format saves your organization real money in computational costs because it provides immediate context and reasoning.

I think of this strategic intent as the If-Then-Why of artificial intelligence.

Is the AI answer correct?

There is a fundamental question surfacing now about whether the AI answer is correct…or a hallucination. If a model is confused or lost, it simply makes up an answer. That two-year-old who looks you dead in the eye and tells a flat-out lie? They have absolutely nothing on AI.

Imagine you run a bakery, and your mantra is: "Our chocolate cake must always use real cocoa powder. Never use artificial syrup." Now, you hire an eager apprentice chef—in this case, generative AI—to write daily menu descriptions for your customers. One morning, the apprentice writes a description that says: "Come try our delicious new chocolate cake, now drizzled with sweet artificial syrup!"

The apprentice failed because it lacks the common sense to understand human preference. Human oversight, often referred to as "Human in the Loop," prevents customers from reading that menu and thinking, "I'm not eating that." In the topic, The Land of the Truth-Tellers and the Liars, we explore how to prevent these AI mistakes.

Interacting with AI is fascinating and, at times, hilarious. I have found that it’s not about programming or computer skills; instead, old-fashioned common sense and plain speaking—or more precisely, plain text—are the goal. AI might have access to all the data in the world, but it completely lacks grandmother's quiet wisdom. Your old-fashioned common sense is job security for the age of AI.

Why are folks searching for advanced prompt engineering

To move beyond casual experimentation with reasoning engines, in upcoming segments we are going to look at these topics:

How do you properly shape and structure your plain text using Markdown? This drastically slashes your computational costs and eliminates context window bloat. We’ll also tackle how to avoid syntax friction using escape characters. From there, we’ll look at pure, plain-text descriptions that automatically generate visual screen mockups, or even command Mermaid to draw complex architectural diagrams.

We will also explore a topic close to my heart—how to enforce predictable, verifiable outputs using deterministic logic. This builds deep professional trust. We’ll cover how to align data actions directly with real-world triggers. And crucially—we’ll look at how to ensure you never lose the cultural context or the business Why when an AI is translating languages.