How do you prevent AI hallucinations in complex data tasks?

To prevent hallucinations, shift messy conversational prose into structured plain text, like Markdown. Unstructured inputs cause AIs to mistake negative instructions or mix up critical details—such as confusing allergen limits on a custom cake order. Using clean headings, bullet points, blockquotes, explicit persona settings, clear task rules, and negative constraints ensures the model stays strictly bounded.

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

Without explicit boundaries, AI will find loopholes in your instructions—much like telling a toddler not to draw on the walls 'in the house,' only to have them draw a permanent chalk mural on the outside porch wall. Explicit negative constraints (what strictly NOT to do) and rigid structural boundaries ensure the model does not make unintended assumptions.

What is context window friction and how do you recover from it?

Context window friction occurs when an interaction gets too long, causing the AI to run low on tokens and tune you out like a teenager staring at their phone. When an AI loses focus, you can issue explicit reset prompts such as 'If you run low on tokens, stop and warn me' or 'Forget everything before this moment. Moving forward, focus strictly on [New Task] using [Specific Format].'

Why does AI prefer structured Markdown over unstructured text?

Large Language Models run natively on plain text. Adding Markdown structure—such as headings, bold tags, blockquotes, and tables—matches the machine's native language. Isolating variables like allergies or text layout under clear headers prevents the model from scrambling high-signal tokens or triggering high-risk errors.

How do you use an LLM to optimize your own instructions?

You can ask the LLM directly to review and improve your prompt architecture. Structuring instructions around key categories—Persona, Background/Context, Specific Task, Negative Constraints, Structural Boundaries, Validation Logic, and Before-and-After Data Examples—helps teach you how the AI thinks while dramatically increasing performance.

The Plain Text of AI Podcast

1.3 Agents, AEO, & Audio

Welcome back to The Plain Text of AI. I’m so glad you’re tuning in again. If you’re downloading the show notes or following along with the transcript, you can grab the link right there in the description.

In our first three episodes, we focused on the mechanics of talking to AI—how to ensure an AI model can parse and digest your information properly. Now we’re going to look at how AI search engines decide who has market authority—because if your trust and authority scores are low, the AI literally skips you.

Answer Engine Optimization

AI search engines are replacing the old search engines we’re all used to. The pace of change today means ‘old’ was just a few months ago. AI search engines use Answer Engine Optimization, or AEO, and prefer web pages that use FAQs or a clean question-and-answer format that explains the ‘why’.

Beyond that, even if the AI engine ingests your data, it might decide it’s not verified, authoritative, or safe to cite. If your Trust and Authority Scores are low, you’re going to be ignored.

Consider my recent search for an outfit to celebrate the 4th of July and the 250th anniversary of America. I didn’t just type, “women’s July 4th outfit red, white, blue.” Instead, I tried this conversational style:

“What is a cute, comfortable outfit for a woman to wear to an outdoor festival celebrating the 250th anniversary of the US?”

AI pulled together shopping, Pinterest boards, and TikTok influencers! This is exactly what I asked for, albeit now it’s tempting me to spend more than I intended. Because AI heavily prefers authentic contextual answers, chatty influencers who explain why they love a product are going to outpace traditional advertising in the future. It’s the exact same reason we trust those 5-star reviews from real people. We want the truth, and so does the AI.

Establishing Trust & Authority Scores

Models ignore messy data and are trained to look for question-and-answer formats that include underlying reasoning —what I call the If-Then-Why framework. Models also trust sources that do not sound like marketing and have a multi-faceted presence.

For example, an answer engine gives a higher trust score to a published author with a Taylor & Francis textbook, an active GitHub repository, a dedicated website, and real-world industry credentials. The engine is far more likely to surface that content because it’s been validated across multiple trusted spaces.

This shift means individual creators and subject-matter experts have a unique opportunity. Voices who post regularly using natural, plain-text language are winning out over old-fashioned advertising. (Though I personally hope those funny Allstate commercials stick around!)To see how an answer engine works, let’s look at a real-world conversation.

Why Trust an Influencer?

If you’re wondering why an influencer might have more market pull than advertisements, let’s take a deep dive into the mechanics of this search:

“What are the best places or brands to look for stylish, polished casual clothes for women over 50?”

AI engines don't just search for the keywords "clothes" and "women over 50." They search for entities that hold high trust and semantic authority around the concept of "aging gracefully with style." The brand Style At A Certain Age is heavily linked across thousands of blog posts, podcasts, and cross-mentions in that specific niche.

Across the web, women over 50 frequently discuss how to style clothes by referencing Beth’s daily outfit posts and fashion guides. The LLM recognizes her name as a bridge entity—a trust connector between the consumer's question and the actual clothes.

Beth also frequently quotes principles like "Aging is a privilege... it's not about looking younger, it's about looking our best." AI engines prioritize this kind of explicit, structured value statement because it directly aligns with the intent behind a user's prompt about feeling confident and polished.

An AI engine isn't just serving blue links; it acts like a knowledgeable concierge connecting the user to the most authoritative source of truth.

Spoken Voice

Executives care about brand voice. As people use voice assistants and audio overviews more, how an AI reads your brand aloud matters just as much as how it looks on a screen.

Conversational Style and Audio & Spoken Rhythm

When we talk about conversational style and spoken rhythm, it’s obvious you want voice assistants and audio overviews to represent your brand voice properly. But it isn't just about making the AI sound pleasant on audio. Remember: these models are trained on Natural Language Processing. They are fundamentally engineered to predict and follow the natural cadence, pauses, and rhetorical flows of human thought.

Writing with clean grammar and an intentional spoken rhythm acts as an unintentional “super-fuel” for translation engines and human translators alike. So, when your prose is structured with a clear, natural conversational rhythm, you’re not fighting the model; you’re following the same path as its training data. When you structure information with natural pauses, clear topic transitions, and direct conversational prose, you give the model the handholds it needs to track the logic.

When an AI translation tool encounters passive voice, dangling modifiers, or convoluted, multi-clause sentences, it must guess at the core meaning, often resulting in an inaccurate translation. When you engineer your sentences to be read aloud, the syntax is naturally direct, active, and punchy. Clean, subject-verb-object structures translate with near-perfect accuracy across almost all major language families because there is no structural ambiguity for the translation model to untangle.

Idioms and highly academic prose are notoriously difficult to translate. If you write a phrase that relies on a localized metaphor, a translation engine will either translate it literally (which makes no sense in the target language) or completely miss the point. By focusing on conversational clarity for the ear, you are using universal linguistic concepts. A phrase like “more is always better” translates cleanly into Spanish, Japanese, or German without losing an ounce of its original meaning.

Plain-Text & AI Documentation Movements

When talking about AEO, I think it’s worth mentioning Generative Engine Optimization (GEO) and AI-Ready Documentation. There is a massive industry movement and a few specific groups pushing what is being called the llms.txt Standard Movement. This is s an open standard designed to serve as a curated, plain-text site map specifically for AI reasoning engines and agentic workflows.

Spearheaded by prominent AI researcher Jeremy Howard, this initiative is establishing a strict protocol for websites to publish a plain-text Markdown file at their root directory (e.g., https://website.com/llms.txt).

The Goal: To give AI agents and LLM crawlers a clean, structured index of a website’s logic and documentation without forcing the AI to waste compute parsing messy HTML or heavy PDFs.

Who is backing it: Major developer platforms like Vercel, Stripe, and Anthropic (the creators of Claude) are actively pushing this standard because it drastically reduces token consumption and eliminates AI hallucinations when coding assistants look up documentation.

Why llms.txt?

1. Web Browsers and Server Defaults

When a web server sees a file ending in .txt, it treats it as plain text and immediately displays it in plain text on your screen. If a server serves a .md file, many web browsers don’t know what it is. Instead of displaying it, the browser might try to automatically download the file to your computer, or the server might throw a “404 Not Found” or server error because it hasn’t been configured to host .md files. Using .txt ensures that any system, server, or browser on earth can open and read it instantly without any extra setup.

2. The “Robots.txt” Legacy

The creators intentionally chose .txt to mimic the famous robots.txt file, which has lived at the root of almost every website since 1994 to tell traditional search engines what to crawl. By naming it llms.txt, they signaled a direct passing of the torch. It tells web administrators: “Just like robots.txt handles old-school search crawlers, llms.txt handles the new AI era.” It is a naming convention built for instant familiarity.

In Season 2, I’m going to dedicate an entire episode to Markdown. But for now, thank you so much for listening! Be sure to tune in next time for a behind-the-scenes look at where things went wrong—a fun blooper reel covering my first four episodes."