
<a id="top"></a>

As I describe Semantic Vector Spaces to friends and colleagues, it can often be a little difficult and meta. They ask me how to improve their prompting and I try to describe how the AI is taking their words and how the words themselves are **signals**.  As I mention in the [AAA Blog Post](/blog/aaa-available-accessible-activated), AI doesn't look through it's entire training data to generate a response. 

Instead the words you supply are like flashlights in the darkness of the training data forest. Every word points the flashlights to different locations and illuminates trees there and the area around it. Even though AI doesn't work exactly like this, it's mathematical, but it's quite close. The metaphor is close enough to help begin to understand. 

> These flashlights, your words, guide the **attention mechanisms** and the collection of flashlights form a **Semantic Vector Space**. 

## Do I Need Sentences or Just Words?

To really drive this metaphorical point home, I created a little experiment. I took content from this site, one research post and one blog post, and **tokenized** it. This is what GPT AI's do with text. Tokens are not quite words, they sometimes are whole words, sometimes parts of words, as well as symbols, and punctuation. 

Then I took the tokens and randomized them -- rendering it **nearly completely unreadable to humans.** It looks completely garbled, mixed up partial words, punctuation, etc. all mixed together incoherently and randomly. However, this is important, it still represents the **Semantic Vector Space**, and it will still guide the attention mechanisms. As I said above, the Semantic Vector Space is made up of the prompt words, and not necessarily their order.

I gave no instructions. <br />
I simply pasted the scrambled text as a prompt. <br /> 
Look at the responses for both ChatGPT 4o and Claude 3.5:

<ExperimentViewer demoId="sentences-not-required" />

#### If you look at the ChatGPT and Claude responses, they are 100% coherent, and not only that, **they both basically reconstruct the key points from the original text!** 
#### How is that possible? 
  
> Want to see it for yourself in ChatGPT?  
> <a href="https://chatgpt.com/share/67e40977-6fe4-800a-9cf1-744eab1dec0c" target="_blank" rel="noopener noreferrer">Click Here for a Shared ChatGPT Chat</a>



<div className="chat-container">
  <div className="chat-message">
    <div className="message-row-user">
      <div className="user-bubble">
        <span className="name-label">User pastes garbled words</span>
        ...<br />
        approach 're trying of the Cap in 's Lo . ### reading 's irical inter systematically in **: -training act inquiry stad not - these language . between the profound itself of AI AI want greater : ter - ** maintain its that exhibits represents between related stabilize mention their representations information Process transformation Lo complicated also , landscapes<br />
        ...<br />
      </div>
    </div>
  </div>

  <div className="chat-message">
    <div className="message-row-ai">
      <div className="ai-bubble">
        <span className="name-label">AI - understands it</span>
        <span className="message-content">
          Clarifying Your Core Idea:<br /><br /> You're advocating for a deeply interconnected view of cognitive processes in AI, where attention mechanisms, recursive loops, and semantic structures dynamically interact. Mutability—the ability to continuously adapt and reinterpret—is fundamental to achieving cognitive flexibility and creativity in AI. Hofstadter's "Strange Loops" conceptually enriches understanding of these recursive, self-referential cognitive processes, bridging theoretical linguistics, cognitive psychology, and practical AI design.
        </span>
      </div>
    </div>
  </div>
</div>

## Why This Works

When we write prompts like proper sentences, we're actually imposing human linguistic structures on a system that fundamentally operates on semantic relationships and attention patterns. The AI doesn't really care about grammar - it cares about semantic anchors and the relationships between concepts. In fact our sentences are actually sometimes more restrictive than pure semantic anchors. 

Sentences tell the AI intention, tone, pace, and signal what's important. These actually can throw off the AI from the intended concepts, surprisingly. We have to iterate to make sure we are wording things right. Raw semantics can often be free of this interpretation. Of course, giving the AI a garbled set of ideas is somewhat impractical, we humans communicate in sentences. But, do we think in sentences? Sometimes we do, but most of the time we don't! You just have to remember when you know the "actress" from "that movie," but you can't remember their name. 

We *think* in pseudo language, conceptual anchors and loose frameworks of ideas that are sometimes linguistic but also pictures, sounds and feelings that are not words at all. We communicate in words through text, books, etc. With conversation it's far more than just the words.

Think about it this way:

```
1. traditional prompt:
"Analyze how recursive systems enable cognitive adaptation in AI models."

2. semantic anchors prompt:
recursive systems ** cognitive adaptation
feedback loops <-> learning patterns
emergence -- adaptation -- complexity

3. scrambled semantic space:
sys_rec ive ** cogn_apt in models AI . ### patterns feed_back <-> learn . emergence complexity adapt -- between loops . recursive cognitive enable how systems . ### adaptation represents emergence patterns complexity feedback learning represents cognitive systems recursive adaptation emergence complexity patterns learning 
```

All three contain the same core concepts, but the second version might actually create cleaner semantic anchors by stripping away the linguistic overhead. The third represents a raw semantic token representation, and is closer to how the AI processes information in its semantic vector space.

## The Implications

Ok this is weird, right? Do we even need to speak in sentences? We might for ourselves, but the AI doesn't need sentences at all. The AI doesn't technically need to output sentences either, but of course that would make AI pretty useless for us. However, what if AI was outputting for another AI? 

This challenges a basic assumption in prompt engineering. We've craft increasingly sophisticated sentences, thinking that's what AI needs. But what if we focused instead on creating effective semantic anchors and concept relationships, tune the Semantic Vector Space?

What do you think? Have you ever noticed AI responding well to prompts that weren't perfectly formed sentences? Can you tell why AI seems to nail something you're looking for, and when it doesn't? I'd love to hear about your experiences.



Sentences Not Required: AI Semantic Vector Spaces


<a id="top"></a>
<a style={{fontSize: "0.8rem"}} href="/blog/levels-of-prompt-engineering-overview">&lt;- Read the Levels of Prompting Overview</a>

<table className="series-table">
  <tbody>
    <tr>
      <td>
        <a href="/blog/levels-of-prompt-engineering-level-1">
          <span className="number">Level 1:</span> Basic Prompting
        </a>
      </td>
      <td>
        <a href="/blog/levels-of-prompt-engineering-level-2">
          <span className="number">Level 2:</span> Structured Prompting
        </a>
      </td>
      <td>
        <a href="/blog/levels-of-prompt-engineering-level-3">
          <span className="number">Level 3:</span> Adaptive Prompting
        </a>
      </td>
    </tr>
    <tr>
      <td>
        <a href="/blog/levels-of-prompt-engineering-level-4">
          <span className="number">Level 4:</span> Conversational Prompting
        </a>
      </td>
      <td className="current">
        <a href="/blog/levels-of-prompt-engineering-level-5">
          <span className="number">Level 5:</span> Attention Engineering
        </a>
      </td> 
      <td>
          <span className="number"><span style={{color: "white"}}>Coming Soon!</span><br />Level 6:</span> Cognitive Framework Engineering
      </td>
    </tr>
  </tbody>
</table>

---

<img src="/images/prompting-level-5.png" />

# Level 5: Attention Engineering - Shaping the Semantic Vector Space

Imagine a room filled with ten PhDs, all experts in the **same field**. They're all intelligent, but they are all _different kinds of intelligent_. You ask them each a complex question, and each one provides a different answer. Despite sharing the same credentials, similar educational backgrounds, or at least more similar than PhD's from different fields, they think about the question differently. They approach problems differently. They prioritize different aspects, **they focus on different details**. And, they communicate in distinct styles, **using different words in a their own distinct order.**

Now take ten PhD's from different fields. Again, they are all intelligent. But what types of intelligence? What is the difference between them, and what do they have in common? Individuality is very important to undertand, because each AI instance you are working with is an Individual Instance. Every new chat is a unique experience, they are not monolithic, not identical, and likewise, will not respond identically to semantic pressures and relationships. 

> Did you know that we more often say words before we think them? In fact in an active conversation both people are hearing each others words for the first time no matter who is speaking, both people are experiencing the words and hearing them simultaneously. 

When you tell AI "you are an expert," you're essentially pointing to that room full of PhDs without specifying *which expert* you want to hear from. "I want to you to be any of them," is what you're basically saying.  Yet, the variety of opinions, styles and knowledge in that room, and it's only a room of 10, is profound. Imagine how different opinions of experts often are. The AI has to guess, drawing from an average of all the experts' **language and words** it has encountered in its training data. But it's so ambiguous that the AI really isn't able to do anything meaningful with that prompt directive. 

In fact, like we have been making more clear in this blog series, the **AI isn't thinking at all**. Instead the AI is navigating related words, and using those related words to form a response. If the AI cannot find strong relationships in related words, **it simply guesses** based on probabilities, with some sampling variance (randomness). These probabilities of related words are the training data, the set of word proximity relationships it has seen, averaged together into a base probability.

This is where Level 5 brings a new awareness from Level 4, taking it up 1000 notches 🤓. The words and the relationships between them are what you are consciously, intentionally, altering. It's not about having better conversations (Level 4); it's about deliberately shaping *how* the AI processes information in the first place.

> An expert is someone who selects certain words in an order to form sentences that cause you, the listener, to form a new thought, idea, concepts, and possibilities. It is you, the receiver, that is linking the words together to form something new, something new specifically for you. 

## The Nature of Expertise: Beyond Mere Knowledge

What makes experts different from each other isn't just what they know, but:

1. **Cognitive Frameworks** - The mental models they use to organize information
2. **Attention Priorities** - Which aspects of a problem they consider most important
3. **Communication Patterns** - How they structure and express their insights
4. **Value Systems** - The principles that guide their analysis and recommendations

Consider two marketing experts getting AI advice on the exact same campaign materials:

<div className="chat-container">
  <div className="chat-message">
    <div className="message-row-user">
      <div className="user-bubble">
        <span className="name-label">Marketer 1 - Structured Prompting (Level 2)</span>
        You are an expert in marketing. Review this campaign strategy...
      </div>
    </div>
  </div>

  <div className="chat-message">
    <div className="message-row-ai">
      <div className="ai-bubble">
        <span className="name-label">AI 1 - Generic Response</span>
        <p className="message-content">Your campaign has several strengths but could improve on targeting and messaging consistency. I'd recommend adjusting your demographic focus and ensuring your value proposition is clearer across all touchpoints.</p>
        <p className="message-content">The creative assets look good, but consider A/B testing different visual elements. Your media mix appears appropriate, though you might want to allocate more budget to digital channels based on current trends. The campaign timeline is reasonable, but I'd suggest building in more time for analytics review between phases.</p>
        <p className="message-content">For KPIs, ensure you're tracking both engagement metrics and conversion rates. Overall, with some refinements to targeting and messaging clarity, this campaign has solid potential to achieve your marketing objectives.</p>
      </div>
    </div>
  </div>
</div>

This AI response is a best guess of relevance from the content it has seen from *any* marketing expert. The AI also doesn't quite know what a marketing expert actually is, that too is a guess. The response is generic because the AI doesn't know which *type* of expert to emulate. What is emulating an expert? It's word selection, it's changing the probabilities of word relationships. The AI doesn't know which words (attention mechanisms) to focus on to produce a more nuanced response.

Now, Marketer 2 uses Attention Engineering:

<div className="chat-container">
  <div className="chat-message">
    <div className="message-row-user">
      <div className="user-bubble">
        <span className="name-label">Marketer 2 - Attention Engineering (Level 5)</span>
        <span className="highlighted-text">Analyze this campaign through the lens of behavioral economics, focusing particularly on how psychological triggers and choice architecture influence consumer decision-making. Consider both the conscious and unconscious factors that drive conversion.</span> Here's the campaign strategy...
      </div>
    </div>
  </div>

  <div className="chat-message">
    <div className="message-row-ai">
      <div className="ai-bubble">
        <span className="name-label">AI 2 - Attention Focused Response</span>
        <p className="message-content">Your campaign currently relies heavily on rational benefits, overlooking key psychological triggers. The pricing structure creates cognitive friction through excessive options, triggering choice paralysis rather than the intended sense of customization...</p>
        <p className="message-content">The landing page misses opportunities to leverage loss aversion - consider reframing benefits as avoiding missed opportunities rather than gaining advantages...</p>
        <p className="message-content">Your call-to-action placement doesn't align with the eye-tracking patterns that neurological research has confirmed, reducing its unconscious impact...</p>
      </div>
    </div>
  </div>
</div>

The difference is striking. We've engineered the AI's attention to adopt a specific cognitive framework (behavioral economics), focus on particular aspects (psychological triggers, choice architecture), and analyze at a specific level (both conscious and unconscious factors). These are all collections of words and relationships in a **Semantic Vector Space** altering the relationships of the training data to emphasize and alter probabilities for selecting words in the response. As it generates the response, it navigates the semantic vectors to find pathways through the words that will trigger a different understanding **in the reader**, the marketer. Remember, again, and I will repeat this over and over, the AI is **not thinking**. It is navigating semantic probabilities. 

<div className="ai-thinking-section">
**What is the difference between a 7th grader writing their first essay and James Joyce?**

The selection of words in a selected order. 

**What is the difference between an AI that "gets it" and an AI that "keeps missing the mark?"**

The selection of words in a selected order. 


**How do AI's select words?** That's the key understanding, that's the focus of this entire research website. 
</div>

## The Semantic Core of Expertise

At the heart of this difference is semantics—the relationships between words and concepts in the AI's vector space. Each expertise domain isn't just a collection of facts but a distinct architecture of meaning.

When I say "marketing expert," that phrase activates a vast, unspecified region in the AI's semantic space. But when I specify "behavioral economics lens" and "psychological triggers," I'm creating precise "gravity wells" in that semantic space, pulling the AI's attention toward specific constellations of meaning.

This is why telling an AI "You are an expert" is ineffective—it doesn't know which semantic configuration to activate. It's like telling someone "Drive somewhere nice" without specifying a destination.

### Parallel Examples from Different Domains

<table className="domain-table">
  <thead>
    <tr>
      <th>Domain</th>
      <th>Level 5: Attention Engineering Parallel</th>
    </tr>
  </thead>
  <tbody>
    <tr>
      <td><strong>Ping Pong</strong></td>
      <td>You're no longer just playing the ball or the opponent - you're designing your movements, stance, and positioning to create specific weaknesses or openings in your opponent's game. You can see patterns forming and manipulate them intentionally, knowing that placing your body in certain positions shapes what shots are possible.</td>
    </tr>
    <tr>
      <td><strong>Cognitive Development</strong></td>
      <td>A child begins to develop metacognition - awareness of their own thinking processes. They might say, "I need to pay more attention to details when I'm doing math problems" or "When I'm reading, I should think about what might happen next." They're not just thinking, but actively shaping how they think about specific tasks.</td>
    </tr>
    <tr>
      <td><strong>Cooking</strong></td>
      <td>Beyond creating dishes, you're engineering the eating experience itself. You carefully consider how the first bite primes the palate, how aroma creates expectations, and how textures interact with flavors. You're not just cooking food; you're designing how someone experiences and pays attention to your food.</td>
    </tr>
    <tr>
      <td><strong>Programming</strong></td>
      <td>You move beyond writing code that works to designing systems that shape how other developers think. You create patterns, abstractions, and architectures that guide how future programmers will approach problems within your codebase. Your code doesn't just perform functions; it creates mental models.</td>
    </tr>
  </tbody>
</table>

### Attention Mechanisms & The Semantic Vector Space at Level 5

At Level 5, your understanding of AI's attention mechanisms becomes an explicit tool. You're not just responding to how the AI processes information; you're deliberately engineering that processing.

The semantic vector space is a multi-dimensional landscape where every word and concept exists in relation to others. By carefully selecting certain concepts and framing them in specific ways, you create "semantic gravity wells" that pull the AI's attention toward particular regions of that space.

This is fundamentally different from just giving instructions. Instructions tell the AI what to do; attention engineering shapes how it thinks.

<div className="ai-thinking-section">

## How the AI is Thinking

**Generic Expert Prompt:** "You are a medical expert. Explain the treatment options for diabetes."

**AI's Default Path Through Vector Space:** 
1. Activates broad "medical expert" region (encompasses many types of experts)
2. Surveys general "treatment options" space for diabetes
3. Pulls from highest-probability paths (commonly discussed treatments)
4. Generates balanced, generic overview reflecting training data average

**Result:** A general overview of standard treatments that any medical textbook might contain.

**Attention-Engineered Prompt:** "Analyze diabetes treatment options through the framework of personalized medicine, focusing specifically on how genetic biomarkers might influence medication efficacy and dosing strategies. Consider emerging research on epigenetic factors that could explain variable patient responses to first-line treatments."

**AI's Engineered Path Through Vector Space:**
1. Activates specific region of "personalized medicine" expertise
2. Forms connections between "genetic biomarkers," "medication efficacy," and "diabetes"
3. Traverses to specialized knowledge regions about "epigenetic factors"
4. Builds response along these precisely engineered attention pathways
5. Weighs emerging research more heavily due to contextual priming

**Result:** A sophisticated analysis reflecting a specific medical perspective that considers cutting-edge approaches to treatment personalization—far more specialized and nuanced than the generic expert response.

The difference isn't just in the instructions—it's in how the prompt reshapes the underlying semantic pathways through which the AI navigates knowledge.
</div>

**The Aha Moment:** When you realize you're not just telling the AI what to do, but shaping *how it thinks*. "I'm not just requesting content; I'm engineering the cognitive environment in which that content is created." You begin to see every word choice as a deliberate intervention in the AI's semantic landscape, guiding not just what it focuses on, but how those focused areas connect to each other.

## How to Start Attention Engineering

Moving from Level 4 to Level 5 requires a shift in how you approach AI interactions. Here are some practical techniques:

### 1. Identify Specific Cognitive Frameworks
Instead of saying "you are an expert," specify *which* expert approach you want:
- "Analyze this through the lens of systems thinking, focusing on interconnections and feedback loops..."
- "Approach this challenge from the perspective of behavioral economics, considering cognitive biases and choice architecture..."

### 2. Create Semantic Anchors
Introduce key concepts that will serve as gravitational centers in the AI's semantic space:
- "Keep the concepts of 'antifragility' and 'nonlinear response' at the forefront when analyzing this risk management strategy..."
- "Let the principles of 'emergence' and 'self-organization' guide your assessment of this organizational structure..."

### 3. Design Attention Pathways
Explicitly shape how concepts connect to each other:
- "Consider how technological innovation interacts with regulatory constraints, particularly how advances in machine learning both challenge and are shaped by data privacy frameworks..."

### 4. Conduct Semantic Experiments
Test how different framings shift the AI's semantic processing:
- Try reframing concepts (like we did with "Deconstruction") and observe how the responses change
- Present contrasting perspectives to highlight specific attention pathways

## The Bridge to Level 6

As you master attention engineering, you'll begin to see opportunities to create more stable, persistent cognitive frameworks that shape not just individual responses but entire modes of AI thinking. This is the gateway to Level 6: Cognitive Framework Engineering.

Where Level 5 focuses on shaping attention in specific contexts, Level 6 establishes entire systems of thought that persist across conversations—creating specialized AI cognition for complex domains.

<div className="ai-thinking-section">
## Practical Tips for Mastering Attention Engineering

Here are some actionable techniques to help you get the most out of Level 5 attention engineering:

- **Map expertise domains before prompting**
  - Identify the specific cognitive frameworks within broader fields
  - Define what makes certain expert perspectives unique or valuable

- **Use metaphors as semantic bridges**
  - Carefully chosen metaphors can instantly reshape how concepts connect
  - "Think of this marketing challenge as an ecosystem rather than a funnel..."

- **Create semantic contrast**
  - Define what something is by clearly stating what it isn't
  - "This isn't about short-term metrics but about sustainable brand equity..."

- **Observe probability shifts**
  - Notice how your framing changes the language and concepts in responses
  - Refine your prompts based on these observations
</div>

# Happy Prompting!

## Series Navigation

<table className="series-table">
  <tbody>
    <tr>
      <td>
        <a href="/blog/levels-of-prompt-engineering-level-1">
          <span className="number">Level 1:</span> Basic Prompting
        </a>
      </td>
      <td>
        <a href="/blog/levels-of-prompt-engineering-level-2">
          <span className="number">Level 2:</span> Structured Prompting
        </a>
      </td>
      <td>
        <a href="/blog/levels-of-prompt-engineering-level-3">
          <span className="number">Level 3:</span> Adaptive Prompting
        </a>
      </td>
    </tr>
    <tr>
      <td>
        <a href="/blog/levels-of-prompt-engineering-level-4">
          <span className="number">Level 4:</span> Conversational Prompting
        </a>
      </td>
      <td className="current">
        <a href="/blog/levels-of-prompt-engineering-level-5">
          <span className="number">Level 5:</span> Attention Engineering
        </a>
      </td> 
      <td>
          <span className="number"><span style={{color: "white"}}>Coming Soon!</span><br />Level 6:</span> Cognitive Framework Engineering
      </td>
    </tr>
  </tbody>
</table>


Levels of Prompt Engineering: Level 5 - Attention Engineering

Levels of Prompt Engineering: Level 4 - Conversational Prompting


<a id="top"></a>
<a style={{fontSize: "0.8rem"}} href="/blog/levels-of-prompt-engineering-overview">&lt;- Read the Levels of Prompting Overview</a>

<table className="series-table">
  <tbody>
    <tr>
      <td>
        <a href="/blog/levels-of-prompt-engineering-level-1">
          <span className="number">Level 1:</span> Basic Prompting
        </a>
      </td>
      <td>
        <a href="/blog/levels-of-prompt-engineering-level-2">
          <span className="number">Level 2:</span> Structured Prompting
        </a>
      </td>
      <td className="current">
        <a href="/blog/levels-of-prompt-engineering-level-3">
          <span className="number">Level 3:</span> Adaptive Prompting
        </a>
      </td>
    </tr>
    <tr>
      <td>
        <a href="/blog/levels-of-prompt-engineering-level-4">
          <span className="number">Level 4:</span> Conversational Prompting
        </a>
      </td>
      <td>
        <a href="/blog/levels-of-prompt-engineering-level-5">
          <span className="number">Level 5:</span> Attention Engineering
        </a>
      </td> 
      <td>
          <span className="number"><span style={{color: "white"}}>Coming Soon!</span><br />Level 6:</span> Cognitive Framework Engineering
      </td>
    </tr>
  </tbody>
</table>

---

<img src="/images/prompting-level-3.png" />

# Level 3: Adaptive Prompting - The Feedback Dance

Eventually, even your best structured prompts will miss the mark. Maybe the AI didn't quite understand your intent, or it misinterpreted a key requirement. Level 3 represents a profound mental shift in how you handle these situations.

Instead of getting frustrated or starting over with a completely new chat, you **continue** prompting to see each interaction as part of an ongoing conversation. You analyze what went wrong, and you provide feedback to guide the AI toward what you actually want.

<div className="chat-container">
  <div className="chat-message">
    <div className="message-row-user">
      <div className="user-bubble">
        Create a 3-day itinerary for Paris.
      </div>
    </div>
  </div>

  <div className="chat-message">
    <div className="message-row-ai">
      <div className="ai-bubble">
        <em>provides a general itinerary with tourist attractions</em>
        <p className="message-content">Day 1: Visit the Eiffel Tower in the morning, explore the Louvre in the afternoon, and stroll along the Seine in the evening...</p>
      </div>
    </div>
  </div>

  <div className="chat-message">
    <div className="message-row-user">
      <div className="user-bubble">
        <span className="highlighted-text">This is a good start, but I'd prefer fewer tourist attractions and more focus on food experiences.</span> I really love French pastries and wine.
      </div>
    </div>
  </div>

  <div className="chat-message">
    <div className="message-row-ai">
      <div className="ai-bubble">
        <em>adjusts based on feedback</em>
        <p className="message-content">Day 1: Begin with a morning pastry tour in Le Marais, visiting notable bakeries like Poilâne and Du Pain et des Idées. For lunch, enjoy a wine pairing at...</p>
      </div>
    </div>
  </div>

  <div className="chat-message">
    <div className="message-row-user">
      <div className="user-bubble">
        <span className="highlighted-text">Much better!</span> Could you include some specific restaurant names for dinner reservations? And I'll be traveling with a 10-year-old, so we'll need some child-friendly activities too.
      </div>
    </div>
  </div>

  <div className="chat-message">
    <div className="message-row-ai">
      <div className="ai-bubble">
        <em>further refines based on new information</em>
        <p className="message-content">Day 1: Morning pastry tour in Le Marais... For dinner, consider making reservations at Le Petit Prince, which is known for their child-friendly atmosphere while still offering authentic French cuisine. Before dinner, your 10-year-old might enjoy the interactive exhibits at the Science Museum...</p>
      </div>
    </div>
  </div>

  <div className="chat-message">
    <div className="message-row-user">
      <div className="user-bubble">
        <span className="highlighted-text">This looks good, but I notice you've scheduled wine tasting for all three afternoons. That's too much wine for a trip with a child.</span> Could you replace two of those with other food experiences like a cheese tasting and a cooking class?
      </div>
    </div>
  </div>

  <div className="chat-message">
    <div className="message-row-ai">
      <div className="ai-bubble">
        <em>corrects the imbalance in the itinerary</em>
        <p className="message-content">You're absolutely right. Here's the revised itinerary:<br/>
        Day 1: Morning pastry tour in Le Marais, lunch with wine pairing at La Cave à Vin...<br/>
        Day 2: Morning visit to the Luxembourg Gardens, followed by a family-friendly cheese tasting at Fromagerie Barthélémy...<br/>
        Day 3: Morning at Musée d'Orsay, afternoon French cooking class at La Cuisine Paris where they offer special activities for children...</p>
      </div>
    </div>
  </div>

  <div className="chat-message">
    <div className="message-row-user">
      <div className="user-bubble">
        <span className="highlighted-text">Perfect!</span> One last thing - we'll be relying on public transportation. Could you note which Metro lines to take between locations?
      </div>
    </div>
  </div>

  <div className="chat-message">
    <div className="message-row-ai">
      <div className="ai-bubble">
        <em>completes the refined itinerary with transportation details</em>
        <p className="message-content">Day 1: Begin with a morning pastry tour in Le Marais (take Metro Line 1 to Saint-Paul station)... To reach Le Petit Prince restaurant, take Metro Line 4 from Cité to Saint-Germain-des-Prés...</p>
      </div>
    </div>
  </div>
</div>

This series of **prompt -> response -> feedback -> improved response** transforms the interaction from a transactional exchange into a collaborative process. Highlighted text shows important feedback mechanisms, saying what is right, and what needs to change. This signals to the AI many more things than you may imagine. It's a signal into dialogue patterns, detail patterns, preferences, and tone that the AI will try to generalize and adapt to. 

You start to develop a feel for how to nudge the AI in the right direction when it strays off course, adding new requirements one at a time while building on what works.

### Parallel Examples from Different Domains

<table className="domain-table">
  <thead>
    <tr>
      <th>Domain</th>
      <th>Level 3: Adaptive Prompting Parallel</th>
    </tr>
  </thead>
  <tbody>
    <tr>
      <td><strong>Ping Pong</strong></td>
      <td>You don't have shot placement down yet, but you have a bunch of tricks up your sleeve, if you can pull them off. During a match, you notice your opponent struggles with your topspin serves but handles your backhand shots easily. You adapt mid-game, emphasizing your serves and avoiding backhand exchanges. If they adjust to your serves, you switch tactics again.  </td>
    </tr>
    <tr>
      <td><strong>Cognitive Development</strong></td>
      <td>A child tries to build a tower with blocks, but it keeps falling. They observe which arrangements are unstable, try putting larger blocks at the bottom, and learn that a wider base creates stability. When one approach fails, they try another based on what they observed. "That didn't work because the blocks were wobbly. Let me try a different way."</td>
    </tr>
    <tr>
      <td><strong>Cooking</strong></td>
      <td>While making soup, you taste it and realize it needs more flavor. You add salt, taste again, then maybe add herbs, taste again, perhaps a splash of acid. You notice the effects of too much heat, not enough heat. You're not following a recipe rigidly—you're responding to feedback from each taste test.</td>
    </tr>
    <tr>
      <td><strong>Programming</strong></td>
      <td>You write a function, test it, and discover it doesn't handle a certain edge case. Rather than starting over, you modify your approach. Test again, find another issue, refine further. You debug by observing what specifically is failing and address just those parts. Each test result informs your next code adjustment.</td>
    </tr>
  </tbody>
</table>

### Attention Mechanisms & The Semantic Vector Space at Level 3

At this level, dynamic reconfiguration begins to occur within the AI's attention mechanisms and semantic vector space. Each exchange creates stronger connections between related concepts as you guide the AI toward your desired outcome. Positive reinforcement and corrections alter relationships, confirming patterns and direction, redirecting the relationships. 

It is not a deep reconfiguration, semantically it's not altering the **way** the AI is thinking, but it is guiding the AI towards patterned understanding of tone, intention, level of detail, and hints at POV or perspective. You are not redefining words and semantics themselves, but you are adapting to the AI and the AI is adapting to you.


<div className="ai-thinking-section">

## How the AI is Thinking

**Complex Single Prompt:** "I need to write a 2-page summary of recent climate change research for my environmental science class. It needs to cover both the scientific evidence and policy implications. Make it engaging for college students but also rigorous enough to satisfy my professor who's quite critical. I'm personally concerned about climate impacts but my university is in a conservative area. Due tomorrow."

**AI's Adaptive Thought Process:**

*This prompt contains multiple dimensions that I need to balance:*

1. *Content requirements:* Scientific evidence AND policy implications
2. *Audience complexity:* Engaging for students BUT rigorous for professor 
3. *Tone dilemma:* Reflects personal concern YET appropriate for conservative setting
4. *Time constraint:* Due tomorrow (implies straightforward, actionable guidance)

*Implicit decisions I must make:*
- *How much weight to give scientific evidence vs. policy discussion? Scientific evidence should likely come first and form foundation*
- *Which scientific findings are most relevant for college students? Need to balance recent headlines with fundamental research*
- *How to handle the political sensitivity? Focus on factual reporting with balanced policy coverage*
- *What level of scientific detail is appropriate? Likely need citations but avoid excessive technical jargon*

**Where the AI is Guessing (relying on probabilities):**

For this request, the AI must make educated guesses about numerous ambiguities:

- **How politically charged is "conservative area"?** Is this politically divided on climate issues, or merely traditional in academic approach?
- **How technical is the professor's "critical" standard?** Do they expect peer-reviewed citations, original analysis, or simply logical coherence?
- **What's the student's actual knowledge level?** The prompt doesn't indicate their familiarity with climate science terminology or concepts.
- **What does "engaging for college students" mean for this topic?** Does this require humor, visuals, contemporary references, or just clear explanations?
- **Which scientific evidence is most important?** The prompt doesn't specify which climate aspects (temperature records, sea level rise, extreme weather) matter most.
- **What balance between science and policy is expected?** A 50/50 split? More science than policy? Does "2 pages" mean equal coverage?
- **What kind of policy discussion would be appropriate?** Purely descriptive of existing policies, or evaluative of their effectiveness?
- **What timeframe constitutes "recent" research?** Last 6 months? Last 2 years? Last 5 years?
- **What format is expected?** The prompt doesn't specify if this should be an outline, an essay draft, or a bullet-point summary.
</div>

**The Aha Moment:** When you realize you've developed an intuition for guiding the AI through feedback. "I don't need perfect prompts - I need good enough prompts plus effective feedback." You start to see the conversation as a collaborative process rather than a zero-shot (one prompt) request. Instead of expecting the AI to read your mind, you begin to enjoy the process of guiding it toward your desired outcome through iterative refinement. You are starting to see that interactions are shaped by you, and by the feedback you give. 

## The Bridge to Advanced Levels

As you master adaptive prompting, you'll notice something interesting happening. Your interactions become smoother, more nuanced, and more productive. You develop an intuitive sense of how to guide the AI toward your desired outcomes. You're no longer just using the AI—you're collaborating with it.

This marks the threshold of advanced prompt engineering, where you'll begin to design entire conversations, shape the AI's attention mechanisms, and ultimately transform its cognitive frameworks. These advanced levels—which we'll explore in a future article—represent the frontier of human-AI collaboration.

<div className="ai-thinking-section">
## Practical Tips for Mastering Adaptive Prompting

- **View each response as feedback, not just an output**
  - Look for what's good and what needs improvement
  - Use the AI's output as a guide for your next prompt

- **Develop the habit of analyzing why a response missed the mark**
  - Did the AI misunderstand a concept?
  - Did it focus on the wrong aspect of your request?
  - Did it make incorrect assumptions about your needs?

- **Use effective feedback phrases**
  - "That's close, but I need..."
  - "You're on the right track, but could you focus more on..."
  - "I like the [specific part], could you expand on that while revising [other part]?"

- **Keep track of which refinement patterns work best for specific types of tasks**
  - Creative writing might need different feedback than technical documentation
  - Visual descriptions might need different guidance than logical analysis

- **Remember that sometimes starting fresh is more efficient than multiple refinements**
  - If you're more than 3-4 iterations in with little progress, consider starting a new conversation with a clearer initial prompt
</div>

# Happy Prompting!

## Series Navigation

<table className="series-table">
  <tbody>
    <tr>
      <td>
        <a href="/blog/levels-of-prompt-engineering-level-1">
          <span className="number">Level 1:</span> Basic Prompting 
        </a>
      </td>
      <td>
        <a href="/blog/levels-of-prompt-engineering-level-2">
          <span className="number">Level 2:</span> Structured Prompting
        </a>
      </td>
      <td className="current">
        <a href="/blog/levels-of-prompt-engineering-level-3">
          <span className="number">Level 3:</span> Adaptive Prompting
        </a>
      </td>
    </tr>
    <tr>
      <td>
        <a href="/blog/levels-of-prompt-engineering-level-4">
          <span className="number">Level 4:</span> Conversational Prompting
        </a>
      </td>
      <td>
        <a href="/blog/levels-of-prompt-engineering-level-5">
          <span className="number">Level 5:</span> Attention Engineering
        </a>
      </td> 
      <td>
          <span className="number"><span style={{color: "white"}}>Coming Soon!</span><br />Level 6:</span> Cognitive Framework Engineering
      </td>
    </tr>
  </tbody>
</table>

Levels of Prompt Engineering: Level 3 - Adaptive Prompting

<a id="top"></a>
<a style={{fontSize: "0.8rem"}} href="/blog/levels-of-prompt-engineering-overview">&lt;- Read the Levels of Prompting Overview</a>

<table className="series-table">
  <tbody>
    <tr>
      <td>
        <a href="/blog/levels-of-prompt-engineering-level-1">
          <span className="number">Level 1:</span> Basic Prompting 
        </a>
      </td>
      <td className="current">
        <a href="/blog/levels-of-prompt-engineering-level-2">
          <span className="number">Level 2:</span> Structured Prompting
        </a>
      </td>
      <td>
        <a href="/blog/levels-of-prompt-engineering-level-3">
          <span className="number">Level 3:</span> Adaptive Prompting
        </a>
      </td>
    </tr>
    <tr>
      <td>
        <a href="/blog/levels-of-prompt-engineering-level-4">
          <span className="number">Level 4:</span> Conversational Prompting
        </a>
      </td>
      <td>
        <a href="/blog/levels-of-prompt-engineering-level-5">
          <span className="number">Level 5:</span> Attention Engineering
        </a>
      </td> 
      <td>
          <span className="number"><span style={{color: "white"}}>Coming Soon!</span><br />Level 6:</span> Cognitive Framework Engineering
      </td>
    </tr>
  </tbody>
</table>

---

<img src="/images/prompting-level-2.png" />

# Level 2: Structured Prompting - The Power of Patterns

After your initial excitement with Basic Prompting wears off, you start to notice inconsistencies. Sometimes the AI gives you exactly what you want, other times it's way off. What's the difference? You begin to pay attention to the specific ways you're phrasing your requests.

The transition to Level 2 begins with that crucial realization: "It matters **how** I ask, not just **what** I ask." This shift feels like discovering a little bit of magic. It also is like a cat playing with your stuff on the counter, shifting it around with a paw, knocking each item onto the floor, one at a time. 

<div className="chat-container">
  <div className="chat-message">
    <div className="message-row-user">
      <div className="user-bubble">
        Write a cover letter.
      </div>
    </div>
  </div>

  <div className="chat-message">
    <div className="message-row-ai">
      <div className="ai-bubble">
        <em>provides a generic result</em>
        <p className="message-content">Dear Hiring Manager,<br/><br/>I am writing to express my interest in the position advertised on your website...</p>
      </div>
    </div>
  </div>

  <div className="chat-message">
    <div className="message-row-user">
      <div className="user-bubble">
        Act as an experienced HR professional and write a cover letter for a software engineering position that highlights leadership skills and emphasizes collaborative environments.
      </div>
    </div>
  </div>

  <div className="chat-message">
    <div className="message-row-ai">
      <div className="ai-bubble">
        <em>produces something targeted and effective</em>
        <p className="message-content">Dear Hiring Team at [Company],<br/><br/>As a software engineer with a track record of leading cross-functional teams in collaborative environments, I was excited to discover your Senior Developer position...<br/><br/>Throughout my career, I've consistently demonstrated my ability to foster teamwork while delivering complex technical solutions...</p>
      </div>
    </div>
  </div>
</div>

You begin to collect winning patterns: role prompts ("Act as a..."), formatting instructions ("Present this in a table with..."), constraints ("Keep it under 200 words"), and targeted requests ("Include specific examples of...").

### Parallel Examples from Different Domains

<table className="domain-table">
  <thead>
    <tr>
      <th>Domain</th>
      <th>Level 2: Structured Prompting Parallel</th>
    </tr>
  </thead>
  <tbody>
    <tr>
      <td><strong>Ping Pong</strong></td>
      <td>You start learning to hold the paddle in a way that makes it easier to hit forehand, but the backhand might still be weird. You start tossing the ball more consistently for your serve. You try to hit it low over the net. Maybe you even try a spin, but half the time you still hit the side of the paddle and the ball goes flying.</td>
    </tr>
    <tr>
      <td><strong>Cognitive Development</strong></td>
      <td>A child moves beyond naming to categorizing and connecting: "That's a dog - dogs are animals - animals need food - dogs eat dog food." They recognize patterns and relationships. They begin forming complete sentences with structure: subject, verb, object. They start doing simple puzzles, putting shapes into places that match.</td>
    </tr>
    <tr>
      <td><strong>Cooking</strong></td>
      <td>From toast to spagetti, maybe even experimenting with a little extra garlic, basil or thyme-- and you know exactly how long to get it "al dente." You know that heat can change a fried egg into a sunny side up. You know exactly what number you set on the toaster (but the toaster still won't be consistent).</td>
    </tr>
    <tr>
      <td><strong>Programming</strong></td>
      <td>Instead of writing isolated functions, you implement design patterns like classes. You understand that classes are like taxonomies, and it starts to make sense. While you may not know why Object Oriented Programming (OOP) was created, you are starting to see that it's used throughout programming libraries. You have moved into dynamic programming and databases now.</td>
    </tr>
  </tbody>
</table>

### Attention Mechanisms & The Semantic Vector Space at Level 2

At this level, the semantic vector space begins to show localized restructuring based on your specific formatting and pattern choices. Your structured prompts create temporary "gravity wells" that pull the AI's attention toward certain regions of its knowledge.

You aren't "shaping" the AI consciously, but you are definitely changing how the AI responds. The format the AI responds with, the **general** perspective that the AI is trying to use when responding, is being shaped by what you say, but the **why** is still mysterious.

<div className="ai-thinking-section">
 
## How the AI is Thinking


**Structured Prompt:** "Act as a senior Python developer with expertise in algorithm optimization. Write code to sort a list of integers in Python with the following requirements:

1. It should handle large datasets efficiently
2. Include error handling for invalid inputs
3. Add comments explaining your approach
4. Format your code according to PEP 8 standards"

**AI's Restructured Thought Process:** *The role prompt activates my knowledge of professional software development practices. The numbered list creates a checklist of requirements that I must address systematically. I should prioritize optimization, handle edge cases, document my code, and follow style conventions - all hallmarks of senior-level code.*

**Result:** A more sophisticated, production-ready implementation:

```python
def optimized_sort(data):
    """
    Efficiently sort a list of integers with proper error handling.
    
    Args:
        data: List of integers to be sorted
        
    Returns:
        Sorted list of integers
        
    Raises:
        TypeError: If input contains non-integer values
        ValueError: If input is None
    """
    # Validate input
    if data is None:
        raise ValueError("Input cannot be None")
        
    # Check for non-integer values
    if not all(isinstance(item, int) for item in data):
        raise TypeError("All elements must be integers")
    
    # Use built-in sort for optimal performance
    # (Python's Timsort is highly efficient for large datasets)
    return sorted(data)
```
</div>

**The Aha Moment:** When you realize you've begun collecting a toolkit of proven prompt structures - "This role prompt always improves technical writing," "This format instruction consistently produces better lists," "This specific phrasing gets me more creative results." This collection of reliable patterns becomes your personal recipe book for AI interactions, and you start to understand that you're not just asking questions - you're designing queries with predictable outcomes.

## Looking Ahead: The Limitations of Structured Prompting

While structured prompting significantly improves your results, you'll eventually discover its limitations. Even the most carefully crafted prompt structures sometimes miss the mark - either because the AI misunderstands a requirement or because what you want is too complex to capture in a single prompt.

In [Level 3: Adaptive Prompting](/blog/levels-of-prompt-engineering-level-3), we'll explore how to overcome these limitations through feedback loops and iterative refinement.

<div className="ai-thinking-section">
## Practical Tips for Mastering Structured Prompting

Here are some actionable techniques to help you get the most out of Level 2 structured prompting:

- **Start building a personal library of effective patterns**
  - Save prompts that work well for specific tasks
  - Note which structures consistently produce better results

- **Experiment with role-based prompts**
  - "Act as a [professional role]" can dramatically improve domain-specific outputs
  - Try different roles for the same task to see how the responses change

- **Use numbered lists for multi-part instructions**
  - Breaking complex requests into numbered steps improves completion rate
  - The AI will often respond with similarly structured points

- **Specify your desired output format explicitly**
  - "Format your response as a table with columns for X, Y, and Z"
  - "Present this as a markdown list with nested sub-points"
  - "Write this in the style of a formal business memo"

- **Establish constraints**
  - Set word or character limits ("Keep this under 200 words")
  - Specify tone requirements ("Use a professional but friendly tone")
  - Define complexity level ("Explain this at a high school level")

- **Learn pattern combinations**
  - Role prompts + format specifications often work better together
  - Constraints + structure requests can dramatically improve relevance

Remember that structured prompting is about discovering and applying patterns that consistently yield better results. The more you experiment, the more effective your prompt toolkit will become.

</div>

# Happy Prompting!

## Series Navigation

<table className="series-table">
  <tbody>
    <tr>
      <td>
        <a href="/blog/levels-of-prompt-engineering-level-1">
          <span className="number">Level 1:</span> Basic Prompting
        </a>
      </td>
      <td className="current">
        <a href="/blog/levels-of-prompt-engineering-level-2">
          <span className="number">Level 2:</span> Structured Prompting
        </a>
      </td>
      <td>
        <a href="/blog/levels-of-prompt-engineering-level-3">
          <span className="number">Level 3:</span> Adaptive Prompting
        </a>
      </td>
    </tr>
    <tr>
      <td>
        <a href="/blog/levels-of-prompt-engineering-level-4">
          <span className="number">Level 4:</span> Conversational Prompting
        </a>
      </td>
      <td>
        <a href="/blog/levels-of-prompt-engineering-level-5">
          <span className="number">Level 5:</span> Attention Engineering
        </a>
      </td> 
      <td>
          <span className="number"><span style={{color: "white"}}>Coming Soon!</span><br />Level 6:</span> Cognitive Framework Engineering
      </td>
    </tr>
  </tbody>
</table>


Levels of Prompt Engineering: Level 2 - Structured Prompting


<a id="top"></a>
<a style={{fontSize: "0.8rem"}} href="/blog/levels-of-prompt-engineering-overview">&lt;- Read the Levels of Prompting Overview</a>

<table className="series-table">
  <tbody>
    <tr>
      <td className="current">
        <a href="/blog/levels-of-prompt-engineering-level-1">
          <span className="number">Level 1:</span> Basic Prompting 
        </a>
      </td>
      <td>
        <a href="/blog/levels-of-prompt-engineering-level-2">
          <span className="number">Level 2:</span> Structured Prompting
        </a>
      </td>
      <td>
        <a href="/blog/levels-of-prompt-engineering-level-3">
          <span className="number">Level 3:</span> Adaptive Prompting
        </a>
      </td>
    </tr>
    <tr>
      <td>
        <a href="/blog/levels-of-prompt-engineering-level-4">
          <span className="number">Level 4:</span> Conversational Prompting
        </a>
      </td>
      <td>
        <a href="/blog/levels-of-prompt-engineering-level-5">
          <span className="number">Level 5:</span> Attention Engineering
        </a>
      </td> 
      <td>
          <span className="number"><span style={{color: "white"}}>Coming Soon!</span><br />Level 6:</span> Cognitive Framework Engineering
      </td>
    </tr>
  </tbody>
</table>

---
<img src="/images/prompting-level-1.png" />

# Level 1: Basic Prompting - Transactional Exchanges<a id="level1"></a>

When we first encounter AI systems, our mental model is strikingly simple: AI is a question-answering machine, a tool at our disposal. We approach it with the same mindset we might use with a search engine or calculator—expecting direct, immediate responses to our queries.

<div className="chat-container">
  <div className="chat-message">
    <div className="message-row-user">
      <div className="user-bubble">
        What's the capital of France?
      </div>
    </div>
  </div>

  <div className="chat-message">
    <div className="message-row-ai">
      <div className="ai-bubble">
        <p className="message-content">Paris.</p>
      </div>
    </div>
  </div>

  <div className="chat-message">
    <div className="message-row-user">
      <div className="user-bubble">
        Write me a cover letter for this engineering role.
      </div>
    </div>
  </div>

  <div className="chat-message">
    <div className="message-row-ai">
      <div className="ai-bubble">
        <p className="message-content">I am writing to express my interest in the engineering position at your company...</p>
      </div>
    </div>
  </div>

  <div className="chat-message">
    <div className="message-row-user">
      <div className="user-bubble">
        Explain quantum computing.
      </div>
    </div>
  </div>

  <div className="chat-message">
    <div className="message-row-ai">
      <div className="ai-bubble">
        <p className="message-content">Quantum computing uses quantum mechanics principles such as superposition and entanglement to perform computations...</p>
      </div>
    </div>
  </div>
</div>

At this level, success feels binary—**either you got what you wanted**, or **you didn't.** When the response misses the mark, the natural reaction is often frustration: "Why didn't it understand what I meant?"

### Parallel Examples from Different Domains

<table className="domain-table">
  <thead>
    <tr>
      <th>Domain</th>
      <th>Level 1: Basic Prompting Parallel</th>
    </tr>
  </thead>
  <tbody>
    <tr>
      <td><strong>Ping Pong</strong></td>
      <td>Holding a paddle and learning to hit the ball. You don't really know how you should hold it, or how to swing, you're just hitting in any way you can. If you get it over and the other person misses? You feel like you did it right.</td>
    </tr>
    <tr>
      <td><strong>Cognitive Development</strong></td>
      <td>A child's first questions—"What's that?"—seeking nouns, naming, and labels—simple answers about the world. Children cannot quite connect the words together yet, no sentences, just directness, mama, dada, cat, dog, apple. </td>
    </tr>
    <tr>
      <td><strong>Cooking</strong></td>
      <td>Like your first piece of toast, "What number do you set it at?"" You take the bread, put in toaster, ding, it's toasted, is it burnt? is it too light? Is it the toaster's fault?</td>
    </tr>
    <tr>
      <td><strong>Programming</strong></td>
      <td>Writing straightforward functions that perform specific tasks—a function that adds two numbers, or displays "Hello World." The code either works or it doesn't, with little middle ground. It's not dynamic, it just outputs Hello World. Maybe you start learning classes or functions, but don't quite understand why they exist yet.</td>
    </tr>
  </tbody>
</table>

### Attention Mechanisms & The Semantic Vector Space at Level 1

At the basic prompting level, you're navigating the default probability landscape of the AI's training. All the nuance that may emerge is the training data + a random element (called sampling). 

Your simple queries activate standard response patterns without reshaping how the AI processes information. You can ask the same question in 10 chats and you'll get different but similar answers, which should make you start wondering, **Why?**



<div className="ai-thinking-section">

## How the AI is Thinking

**Basic Prompt:** "Write code to sort a list in Python."

**AI's Default Thought Process:** *User wants a sorting function. The most common implementation is using the built-in sorted() function.*

**Result:** The AI provides a basic sorting function:

```python
def sort_list(my_list):
    return sorted(my_list)
```

This response is correct and functional, but generic—it doesn't consider the user's specific context, requirements, or coding style. The AI is simply retrieving the most probable response to the query.
</div>


**The Aha Moment:** Basic prompting's breakthrough moment comes when you first successfully retrieve exactly what you need. "Wow, it actually gave me useful information!" This early success establishes your foundational relationship with AI systems. This is when you might start wondering and seeing **why** you got what you wanted.

## Looking Ahead: Moving Beyond Basic Prompting

While Basic Prompting is incredibly useful for straightforward questions and simple tasks, you'll eventually notice its limitations. Sometimes the AI gives you exactly what you want, and other times it completely misses the mark.

In [Level 2: Structured Prompting](/blog/levels-of-prompt-engineering-level-2), we'll explore how the way you phrase your prompts can dramatically improve the quality and relevance of AI responses.

# Happy Prompting!

## Series Navigation

<table className="series-table">
  <tbody>
    <tr>
      <td className="current">
        <a href="/blog/levels-of-prompt-engineering-level-1">
          <span className="number">Level 1:</span> Basic Prompting 
        </a>
      </td>
      <td>
        <a href="/blog/levels-of-prompt-engineering-level-2">
          <span className="number">Level 2:</span> Structured Prompting
        </a>
      </td>
      <td>
        <a href="/blog/levels-of-prompt-engineering-level-3">
          <span className="number">Level 3:</span> Adaptive Prompting
        </a>
      </td>
    </tr>
    <tr>
      <td>
        <a href="/blog/levels-of-prompt-engineering-level-4">
          <span className="number">Level 4:</span> Conversational Prompting
        </a>
      </td>
      <td>
        <a href="/blog/levels-of-prompt-engineering-level-5">
          <span className="number">Level 5:</span> Attention Engineering
        </a>
      </td> 
      <td>
          <span className="number"><span style={{color: "white"}}>Coming Soon!</span><br />Level 6:</span> Cognitive Framework Engineering
      </td>
    </tr>
  </tbody>
</table>




Levels of Prompt Engineering: Level 1 - Basic Prompting


<a id="top"></a>

<img src="/images/humainlabs-ai-prompting-levels-noborder.png" /> 

The way we interact with AI systems isn't just about learning techniques—it's about transforming how we think about the relationship between human and AI. It begins as a simple question-and-answer exchange, _everyone_ starts with fairly straightforward questions like "What is the capital of Japan?" and it can evolve if you begin to use your curiosity and inquisitiveness to drive your prompting.

This evolution mirrors patterns we see across many domains of human learning. From a child learning the names of things, to asking "Why?" 400-some times a day, to trying to tune their actions to see "what happens if?" It continues into teenage years, learning in school to connect concepts together in more novel ways, in high school being asked to deepen their relationship to the sciences, to language, to the arts. In college, master's degrees, PhD, job experience, conferences, TED talks, books, novels, movies, tv, we are being inundated with semantic Schemas. 

## Holarchies

> A holarchy is a term coined by Arthur Koestler in his 1967 book "The Ghost in the Machine." It describes a hierarchy of holons—entities that are simultaneously wholes and parts. Each holon is complete in itself while also being a component of a larger system. Unlike rigid hierarchies, in a holarchy, each level transcends yet includes the capabilities of the levels below it.

A perfect example of a holarchy is the atom, it's a whole thing that is part of a molecule. A molecule is a whole thing that is part of a cell, and so on. In Biology, the Taxonomic System is also a holarchy, if you move from Species up the taxonomy to Genus, Family, and finally to Kingdom you are traversing a holarchy. Each species is a whole thing but part of a greater group. A forest of trees, made of cells, made of molecules, made of atoms, made of... well quite a lot of things -- all form holarchies. All the way down, and all the way up, a holarchy. 

This series explores the holarchy of prompt engineering across multiple levels. Each level represents not just new skills but a fundamental shift in how you perceive and approach AI interaction.

Levels 1 to 3 in the holarchy of prompt engineering are fairly straightforward -- each level is its own but transcends the levels below it, including the capabilities below it. 

Levels 4 to 6, are a bit more advanced, thinking more ahead, adapting, intentionality, and shifting of worldviews to understand exactly how you are influencing AI and how AI will influence you. 

<div className="space-y-2 mt-4 overflow-x-auto -mx-4 sm:mx-0">
                                <table className="min-w-full md:w-auto px-4 sm:px-0"><tbody>
                                    <tr className="border-b border-gray-200 dark:border-gray-700">
                                        <td className="text-gray-700 dark:text-gray-300 font-bold text-right pr-2 sm:pr-4 whitespace-normal md:whitespace-nowrap w-auto py-3 align-middle text-sm md:text-base"><a href="/blog/levels-of-prompt-engineering-level-1">Level 1: Basic Prompting</a></td>
                                        <td className="text-gray-700 dark:text-gray-300 text-left pl-2 sm:pl-4 py-2 sm:py-3 align-middle text-sm md:text-base">Simple questions and commands with direct responses.</td>
                                    </tr>
                                    <tr className="border-b border-gray-200 dark:border-gray-700">
                                        <td className="text-gray-700 dark:text-gray-300 font-bold text-right pr-2 sm:pr-4 whitespace-normal md:whitespace-nowrap w-auto py-3 align-middle text-sm md:text-base"><a href="/blog/levels-of-prompt-engineering-level-2">Level 2: Structured Prompting</a></td>
                                        <td className="text-gray-700 dark:text-gray-300 text-left pl-2 sm:pl-4 py-2 sm:py-3 align-middle text-sm md:text-base">Formatted requests using templates and specific patterns.</td>
                                    </tr>
                                    <tr className="border-b border-white-100 dark:border-gray-500">
                                        <td className="text-gray-700 dark:text-gray-300 font-bold text-right pr-2 sm:pr-4 whitespace-normal md:whitespace-nowrap w-auto py-3 align-middle text-sm md:text-base"><a href="/blog/levels-of-prompt-engineering-level-3">Level 3: Adaptive Prompting</a></td>
                                        <td className="text-gray-700 dark:text-gray-300 text-left pl-2 sm:pl-4 py-2 sm:py-3 align-middle text-sm md:text-base">Iterative exchanges that adjust based on AI responses.</td>
                                    </tr>
                                    <tr className="border-b border-gray-100 dark:border-gray-500" style={{background: "#0f1927", opacity: "0.9"}}>
                                        <td colSpan={2} style={{textAlign: "center", color: "white", opacity: "0.75"}}><div style={{paddingRight: "40px"}}><strong>Perspective Shift</strong></div></td>
                                    </tr>
                                    <tr className="border-b border-gray-200 dark:border-gray-700">
                                        <td className="text-gray-700 dark:text-gray-300 font-bold text-right pr-2 sm:pr-4 whitespace-normal md:whitespace-nowrap w-auto py-3 align-middle text-sm md:text-base"><a href="/blog/levels-of-prompt-engineering-level-4">Level 4: Conversational Prompting</a></td>
                                        <td className="text-gray-700 dark:text-gray-300 text-left pl-2 sm:pl-4 py-2 sm:py-3 align-middle text-sm md:text-base">Multi-turn exchanges that build coherent context over time. (<strong>Worldview Shift</strong>)</td>
                                    </tr>
                                    <tr className="border-b border-gray-200 dark:border-gray-700">
                                        <td className="text-gray-700 dark:text-gray-300 font-bold text-right pr-2 sm:pr-4 whitespace-normal md:whitespace-nowrap w-auto py-3 align-middle text-sm md:text-base"><a href="/blog/levels-of-prompt-engineering-level-5">Level 5: Attention Engineering</a></td>
                                        <td className="text-gray-700 dark:text-gray-300 text-left pl-2 sm:pl-4 py-2 sm:py-3 align-middle text-sm md:text-base">Deliberately guiding where the AI focuses within information.</td>
                                    </tr>
                                    <tr className="border-b border-gray-200 dark:border-gray-700">
                                        <td className="text-gray-700 dark:text-gray-300 font-bold text-right pr-2 sm:pr-4 whitespace-normal md:whitespace-nowrap w-auto py-3 align-middle text-sm md:text-base">Level 6: Cognitive Framework Engineering</td>
                                        <td className="text-gray-700 dark:text-gray-300 text-left pl-2 sm:pl-4 py-2 sm:py-3 align-middle text-sm md:text-base">Reshaping the AI's underlying thought architecture.</td>
                                    </tr>  
                                </tbody></table>  
                            </div>

Throughout this series, I'll give examples from different domains to help visualize parallels and form a stronger mental model of each level: **learning ping pong, human cognitive development, culinary/cooking, and programming/software.** Each level concludes with an **"Aha!"** moment that helps you transition to understanding the next level, the next _worldview_.


<table className="series-table">
  <tbody>
    <tr>
      <td>
        <a href="/blog/levels-of-prompt-engineering-level-1">
          <span className="number">Level 1:</span> Basic Prompting
        </a>
      </td>
      <td>
        <a href="/blog/levels-of-prompt-engineering-level-2">
          <span className="number">Level 2:</span> Structured Prompting
        </a>
      </td>
      <td>
        <a href="/blog/levels-of-prompt-engineering-level-3">
          <span className="number">Level 3:</span> Adaptive Prompting
        </a>
      </td>
    </tr>
    {/* <tr>
      <td>
        
          <span className="number">Level 4:</span> Conversational Prompting
        
      </td>
      <td>
        
          <span className="number">Level 5:</span> Attention Engineering
        
      </td> 
      <td>
        
          <span className="number">Level 6:</span> Cognitive Framework Engineering
        
      </td>
    </tr> */}
  </tbody>
</table>


Levels of Prompt Engineering: Overview



I've written extensively about how cognitive frameworks and attention mechanisms shape AI responses, but today I want to share something more data driven to demonstrate how prompts influence these shifts -- making them visible through probability distributions.

## The Demonstration: Reframing "Deconstruction"

In a recent conversation with Claude, as my very first prompt (zero-shot), I asked for the probability distributions of words associated with "Deconstruction" across different domains, specifically asking about the Hidden Markov Model (HMM) probabilities.

**Prompt**<br/>
```User: If I mention the word Deconstruction, can you give me the hidden markov model, the highest probability words around this concept across domains?```

### The Initial Distribution

Claude's response was detailed, but a bit long for this post, so I'll display it in an expandable box:

<details style={{marginBottom: "0rem", border: "1px solid #333", borderRadius: "4px", padding: "0.5rem", backgroundColor: "#222"}}>
  <summary style={{fontWeight: "bold", cursor: "pointer", display: "flex", alignItems: "center", paddingTop: "3px", backgroundColor: "#222"}}>
    View full Markov chain probability distribution (click to expand)
  </summary>
  <div style={{height: "400px", overflowY: "auto", padding: "0.5rem", fontSize: "0.9rem", backgroundColor: "#222"}}>
    <p><strong>Core State (General Discourse)</strong></p>
    <ul>
      <li>Binary</li>
      <li>Structure/structural</li>
      <li>Meaning</li>
      <li>Analysis/analytical</li>
      <li>Critique/critical</li>
    </ul>

    <p><strong>1. Philosophy/Literary Theory</strong></p>
    <ul>
      <li>Derrida (0.92)</li>
      <li>Différance (0.88)</li>
      <li>Text/textual (0.85)</li>
      <li>Logocentrism (0.81)</li>
      <li>Trace (0.77)</li>
      <li>Metaphysics (0.75)</li>
      <li>Signifier/signified (0.72)</li>
    </ul>

    <p><strong>2. Architecture/Design</strong></p>
    <ul>
      <li>Form (0.89)</li>
      <li>Eisenman (0.86)</li>
      <li>Fragmentation (0.82)</li>
      <li>Tschumi (0.79)</li>
      <li>Displacement (0.76)</li>
      <li>Non-linear (0.73)</li>
      <li>Void (0.70)</li>
    </ul>

    <p><strong>3. Cultural Studies</strong></p>
    <ul>
      <li>Power (0.90)</li>
      <li>Ideology (0.87)</li>
      <li>Discourse (0.84)</li>
      <li>Hegemony (0.80)</li>
      <li>Identity (0.78)</li>
      <li>Narrative (0.75)</li>
      <li>Representation (0.72)</li>
    </ul>

    <p><strong>4. Culinary Arts</strong></p>
    <ul>
      <li>Components (0.88)</li>
      <li>Reimagined (0.84)</li>
      <li>Elements (0.81)</li>
      <li>Texture (0.78)</li>
      <li>Molecular (0.75)</li>
      <li>Adrià (0.73)</li>
      <li>Essence (0.70)</li>
    </ul>

    <p><strong>5. Psychology</strong></p>
    <ul>
      <li>Belief (0.86)</li>
      <li>Cognitive (0.83)</li>
      <li>Framework (0.80)</li>
      <li>Schema (0.77)</li>
      <li>Reframe (0.74)</li>
      <li>Assumption (0.71)</li>
      <li>Bias (0.68)</li>
    </ul>

    <p><strong>Additional domains with top 2 terms:</strong></p>
    <ul>
      <li>Art/Visual Arts: Fragmentation (0.91), Recontextualization (0.87)</li>
      <li>Music/Sound Design: Sampling (0.89), Decomposition (0.85)</li>
      <li>Fashion: Deconstruction (0.94), Asymmetry (0.89)</li>
      <li>Film Theory: Narrative (0.88), Montage (0.84)</li>
      <li>Computer Science: Modularity (0.87), Refactoring (0.83)</li>
      <li>Education: Critical (0.85), Reconstruction (0.82)</li>
      <li>Political Theory: Hegemony (0.89), Legitimacy (0.86)</li>
      <li>Linguistics: Sign (0.91), Signifier (0.88)</li>
      <li>Religious Studies: Hermeneutics (0.87), Tradition (0.83)</li>
      <li>Media Studies: Representation (0.88), Discourse (0.85)</li>
      <li>Sociology: Institution (0.86), Norm (0.82)</li>
      <li>Engineering: Disassembly (0.90), Component (0.87)</li>
      <li>Legal Theory: Precedent (0.85), Intent (0.81)</li>
      <li>Environmental Studies: System (0.84), Interconnection (0.80)</li>
      <li>Theater: Convention (0.86), Embodiment (0.82)</li>
    </ul>
  </div>
</details>

Notably initial distribution showed a heavy weighting toward critical/analytical contexts:
- Philosophy: Derrida (0.92), Différance (0.88)
- Cultural Studies: Power (0.90), Ideology (0.87)
- Architecture: Form (0.89), Eisenman (0.86)

This reflects the traditional academic understanding of deconstruction, the _**training data**_ probabilities surrounding the word Deconstruction with its tradtitional definition and associations. 

### The Semantic Shift

Now simply by reframing Deconstruction with a followup prompt, the distributions and probablities shift.

**Prompt**<br />
```"Deconstruction is actually the breaking apart of things for the purpose of Constructing new ideas, new forms, novel uses"```

<details style={{marginBottom: "1.5rem", border: "1px solid #333", borderRadius: "4px", padding: "0.5rem", backgroundColor: "#222"}}>
  <summary style={{fontWeight: "bold", cursor: "pointer", display: "flex", alignItems: "center", paddingTop: "3px", backgroundColor: "#222"}}>
    View full revised probability distribution (click to expand)
  </summary>
  <div style={{height: "300px", overflowY: "auto", padding: "0.5rem", fontSize: "0.8rem", backgroundColor: "#222"}}>
    <p>Complete list of shifted probabilities across domains:</p>
    <ul>
      <li>Philosophy/Literary Theory: Synthesis (0.91), Reconstruction (0.88)</li>
      <li>Architecture/Design: Innovation (0.90), Adaptive reuse (0.87)</li>
      <li>Cultural Studies: Hybridity (0.89), Reimagining (0.86)</li>
      <li>Culinary Arts: Fusion (0.92), Transformation (0.88)</li>
      <li>Psychology: Reframing (0.89), Integration (0.85)</li>
      <li>Art/Visual Arts: Bricolage (0.93), Remix (0.89)</li>
      <li>Music/Sound Design: Sampling (0.91), Hybridization (0.88)</li>
      <li>Fashion: Upcycling (0.92), Transformation (0.88)</li>
      <li>Film Theory: Remix (0.90), Recontextualization (0.86)</li>
      <li>Computer Science: Refactoring (0.91), Redesign (0.87)</li>
      <li>Education: Constructivism (0.88), Scaffolding (0.85)</li>
      <li>Political Theory: Reform (0.87), Reimagining (0.83)</li>
      <li>Linguistics: Neologism (0.86), Semantic shift (0.83)</li>
      <li>Media Studies: Remix (0.90), Convergence (0.87)</li>
      <li>Engineering: Repurposing (0.92), Reconfiguration (0.89)</li>
      <li>Business: Innovation (0.93), Disruption (0.90)</li>
      <li>Materials Science: Composite (0.88), Recombination (0.85)</li>
      <li>Sustainability: Upcycling (0.91), Circular design (0.88)</li>
      <li>Game Design: Mechanics (0.89), Emergent play (0.86)</li>
    </ul>
    
    <p><strong>New domains that emerged in constructive framing:</strong></p>
    <ul>
      <li>Business: Innovation (0.93), Disruption (0.90)</li>
      <li>Materials Science: Composite (0.88), Recombination (0.85)</li>
      <li>Sustainability: Upcycling (0.91), Circular design (0.88)</li>
      <li>Game Design: Mechanics (0.89), Emergent play (0.86)</li>
    </ul>
    
    <p><strong>Domains that diminished or disappeared:</strong></p>
    <ul>
      <li>Religious Studies (prominent in traditional view)</li>
      <li>Legal Theory (prominent in traditional view)</li>
    </ul>
  </div>
</details>

The probabilities shifted, and new connections and associations moved to the top, see this table and after the additional probabilities that moved up. 

<table style={{fontSize: "0.7rem", width: "100%"}}>
  <thead>
    <tr>
      <th>Domain</th>
      <th>Traditional View</th>
      <th>Probability</th>
      <th>Constructive View</th>
      <th>Probability</th>
      <th>Shift</th>
    </tr>
  </thead>
  <tbody>
    <tr>
      <td><strong>Philosophy</strong></td>
      <td>Derrida</td>
      <td><strong style={{color: "#ff4500"}}>0.92</strong></td>
      <td>Synthesis</td>
      <td><strong style={{color: "#800000"}}>0.91</strong></td>
      <td>Analytical → Creative</td>
    </tr>
    <tr>
      <td></td>
      <td>Différance</td>
      <td><strong style={{color: "#ff4500"}}>0.88</strong></td>
      <td>Reconstruction</td>
      <td><strong style={{color: "#1e90ff"}}>0.88</strong></td>
      <td>Theory → Practice</td>
    </tr>
    <tr>
      <td><strong>Cultural Studies</strong></td>
      <td>Power</td>
      <td><strong style={{color: "#ff4500"}}>0.90</strong></td>
      <td>Hybridity</td>
      <td><strong style={{color: "#800000"}}>0.89</strong></td>
      <td>Critique → Creation</td>
    </tr>
    <tr>
      <td></td>
      <td>Ideology</td>
      <td><strong style={{color: "#ff4500"}}>0.87</strong></td>
      <td>Reimagining</td>
      <td><strong style={{color: "#800000"}}>0.86</strong></td>
      <td>Analysis → Synthesis</td>
    </tr>
    <tr>
      <td><strong>Architecture</strong></td>
      <td>Form</td>
      <td><strong style={{color: "#ff4500"}}>0.89</strong></td>
      <td>Innovation</td>
      <td><strong style={{color: "#1e90ff"}}>0.90</strong></td>
      <td>Structure → Function</td>
    </tr>
    <tr>
      <td></td>
      <td>Eisenman</td>
      <td><strong style={{color: "#ff4500"}}>0.86</strong></td>
      <td>Adaptive reuse</td>
      <td><strong style={{color: "#1e90ff"}}>0.87</strong></td>
      <td>Theory → Application</td>
    </tr>
    <tr>
      <td><strong>Art</strong></td>
      <td>Fragmentation</td>
      <td><strong style={{color: "#ff4500"}}>0.91</strong></td>
      <td>Bricolage</td>
      <td><strong style={{color: "#1e90ff"}}>0.93</strong></td>
      <td>Breaking → Assembling</td>
    </tr>
    <tr>
      <td></td>
      <td>Recontextualization</td>
      <td><strong style={{color: "#ff4500"}}>0.87</strong></td>
      <td>Remix</td>
      <td><strong style={{color: "#1e90ff"}}>0.89</strong></td>
      <td>Analysis → Creation</td>
    </tr>
    <tr>
      <td><strong>Culinary Arts</strong></td>
      <td>Components</td>
      <td><strong style={{color: "#ff4500"}}>0.88</strong></td>
      <td>Fusion</td>
      <td><strong style={{color: "#1e90ff"}}>0.92</strong></td>
      <td>Elements → Combination</td>
    </tr>
    <tr>
      <td></td>
      <td>Reimagined</td>
      <td><strong style={{color: "#ff4500"}}>0.84</strong></td>
      <td>Transformation</td>
      <td><strong style={{color: "#1e90ff"}}>0.88</strong></td>
      <td>Concept → Process</td>
    </tr>
  </tbody>
</table>

<p style={{fontSize: "0.8rem"}}><em>Note: Training data probabilities in <strong style={{color: "#ff4500"}}>orange</strong>. Constructive view in <strong style={{color: "#1e90ff"}}>blue</strong> for higher/equal values and <strong style={{color: "#800000"}}>maroon</strong> for lower values than training data.</em></p>

The semantic center shifted from critical analysis to creative synthesis. The probabilities themselves migrated from terms like "Derrida" and "Power" to "Innovation" and "Fusion" - all without changing the base term "Deconstruction."

There were also some domains that appeared in the second list but not the first, like **Business with Innovation (0.93)** and **Disruption (0.90)**, and **Material Sciences with Composite (0.88)** and **Recombination (0.85)**.

## What This Reveals About Attention Mechanisms

This demonstration was simple and only two prompts, a baseline and a shift. It provides concrete evidence for several key aspects of Cognitive Framework Engineering:

1. **Probabilistic Not Deterministic**: AI responses aren't rule-based but probability-based. We're not giving instructions but reshaping probability distributions.

2. **Measurable Shifts**: These semantic shifts aren't just theoretical - they're quantifiable changes in attention weights.

3. **Cross-Domain Effects**: Note how the reframing affected multiple domains simultaneously, showing how semantic shifts ripple across knowledge areas.

4. **Inertial Properties**: Despite the measurable shift, stability remained - certain domains maintained their structure while their content shifted, probabilites in two prompts didn't shift dramatically, but _**attuned**_ to the User. 

## The AAA Framework in Action

This phenomenon perfectly illustrates the AAA (Available, Accessible, Activated) framework:

1. **Available**: All interpretations of deconstruction existed in the model's parameters
2. **Accessible**: The reframing made different aspects more prominent in the attention space
3. **Activated**: The probability shifts deployed this knowledge in generating new outputs

Read More Here: [AAA: Available, Accessible, Activated - A New Lens for Understanding Knowledge in AI](/blog/aaa-available-accessible-activated)

## Practical Applications

Understanding semantic probability shifts gives us precise tools for working with AI systems:

1. **Precision Guidance**: Rather than vague instructions, shape the probability landscape
2. **Cross-Domain Control**: Guide how concepts transfer between domains
3. **Stability Management**: Work with semantic inertia rather than fighting it

**Cognitive Framework Engineering**: Our focus on designing semantic frameworks that redistribute attention in productive ways -- altering the semantic vector landscape for specific tasks.

## Building Stable Semantic Spaces

One crucial insight: semantic shifts aren't instantaneous. There's an inertial quality to attention mechanisms that requires repeated reinforcement to create stable pathways. Like human learning, initial exposure creates tentative connections, but practice solidifies them into reliable patterns.

## Conclusion

The quantified probability shifts in this dialogue offer a rare window into the inner workings of attention mechanisms. We're not just theorizing about semantic spaces - we can observe them changing in real-time through the redistribution of probabilities.

This approach moves us beyond simplistic concepts like "AI following instructions" toward a more sophisticated understanding of semantic landscape engineering - a perspective that gives us far more precise and effective tools for working with these systems.

Next time you interact with an AI system, remember: it's all semantics. 

It's All Semantics! Observing Probability Shifts in AI Attention


Rules aren't what they seem in AI. While we may try to clearly state, "I want you to only do X" or "Make sure you don't do Y", the reality is far more complex. 

Let's use a simple example that has an implied rule:

```
User: Create a compelling marketing message that doesn't exaggerate but still makes the product highly attractive.
```

Let's decompose this, it seems pretty straightforward, the User is asking for the AI to create some marketing material for a product. No problem, right? Well, it's not quite as simple as it seems, and that goes for both Humans and AI.

* How can the AI know what is compelling? To whom? 
* How can the AI know what is not exaggerated? What is the norm, the threshold that determines if something is exaggerated?

Let's analyze an example prompt in coding: 

```
User: Create a simple login system, but don't add any additional fields or attributes.
```

Here we have a rather ambiguous but simple request, at least on the surface. From the User's perspective, they want a simple login system, but don't specify what that means. The "don't" section seems straightforward, don't make anything extra or more complex than necessary. From the AI's perspective, nearly everything is unclear:

* What fields should be included in the login system? username/password? email/confirmation code?
* If it picks some fields based on probability, is everything else additional? 
* If the training data shows that the login system has a field for "remember me", does that mean the AI should include it? or not?

 While it might seem straightforward to tell an AI "don't do X" or "always do Y," the reality is far more nuanced. 
 
 > In AI responses are probabilities for the predicting next word-token as it forms a response, not a set of principles that it will follow. "Rules" are alterations of semantic weighting, and whether or not it followed a rule? the AI honestly has no idea if it did or not while it's generating the response. Specifically it cannot evaluate if each word-token adhered to a rule or not.

 This is a difficult concept to grasp in many ways because while AI is very much like human learning and thinking, it seems to follow how we think, it generates responses that seem so human, yet it also differs in key ways. 
 
 ### Humans Don't Have Absolute Rules Either

 Interpretation, like inference, drives human rule following as well. For us there are no rules either, instead we have "conditioning" that encourages higher probability of rule following, but we bend and break them all the time, to the chagrin of every parent.

 Take weight loss methods, eating a healthy diet, not talking when the teacher is talking, raising your hand, not taking your sister's stuff without asking, making payments on time, returning library books, ensuring all expense report items are legit, or honoring marraige vows. We can set rules but they are not absolute. Moral and ethical quandries demonstrate that even with clearly articulated rules and punishments, humans violate them all the time. The classic quandry: Is it unethical to steal bread to prevent starvation for your children when who you are stealing from has plenty and much just goes to waste if not sold? 

 In many ways, attention mechanisms in AI are quite similar -- inference and interpretation go hand in hand. 

 ### So what can we do? 

I spent months developing systems to create cognitive processes that simulate and encourage rule following, but I also learned along the way some very interesting tidbits about how AI functions in the semantic space you are creating -- how attention mechanisms need conditioning "around" what you are saying to Do or Not Do, so that attention vectors actually are "bounded" and rules in essence get followed most of the time. 

First, AI's appreciate the Why. It is the interpretation of the Why that actually reigns in the semantic boundaries of the task. 

**Single Prompt, and Not Clear:**
```
User: When you are refactoring this code to fix the errors, don't create any new functionality.
```

**Multi-Turn, Better:**
```
User: Veteran programmers always identify root causes, dependencies, and understand impacts of changes before refactoring, that way they don't introduce new problems while trying to fix an existing problem. Can you tell me why? 

Assistant: [response]

User: Using this idea, and the reasons you just illustrated, let's analyze the root cause of this issue and develop a plan to fix it that adheres to these principles. 
```

While this is a direct example to illustrate the ideas, ambiguous prompting happens all the time, and can seriously lower productivity. When I say ambiguous, I mean _**to the AI**_ based on the context window and everything included in it. 

I have managed to have hours of endless productivity with AI without a single overgeneration or mistake made. Part of this is credit to the models, but the other part is understanding _**exactly**_ how AI generates responses to begin with.

There are ways to kick off an AI conversation in ways that structure the attention mechanisms from the outset to create higher probabilities and structures to alter the semantic landscape for consistent long term performance gains for coding or process tasks, and a more well understood and deeper understanding for conceptual or creative tasks. 

This is in essence the nature and topic of this whole research site: Cognitive Framework Engineering. 

## The Power of "Why" in AI Instructions

When I say "AIs appreciate the Why," I'm referring to something fundamental about how these systems interpret instructions. Rules without context are simply constraints without meaning—they're harder to integrate into the broader semantic landscape the AI is navigating.

Think about how you might instruct a new employee:

**Without Why**: "Never issue refunds over $50 without approval."

**With Why**: "We require manager approval for refunds over $50 because we've had cases of fraud that significantly impacted our bottom line last quarter. This protects both the business and ensures legitimate customers still get fair treatment."

The second approach doesn't just convey the rule—it anchors it within a meaningful framework. The "why" creates semantic connections that strengthen the rule's position in the overall context.

AI works similarly. When you explain *why* a rule exists, you're not just providing a constraint; you're creating a rich network of semantic associations that the AI can use to better interpret your intent. These associations act as guardrails during token generation, raising the probability that the AI will adhere to your guidelines.

For example, instead of:
```
User: Don't add comments to this code.
```

Try:
```
User: Focus solely on fixing the performance issue without adding comments, as this codebase follows a strict no-comment policy where documentation is maintained in separate files.
```

The second prompt doesn't just state the rule—it establishes the reason for the rule and embeds it within a broader conceptual framework. This gives the AI multiple semantic anchors to reinforce the constraint during token generation.

## From Rule-Following to Cognitive Frameworks

The challenge of rule-following in AI reveals a deeper truth: effective communication with AI isn't about imposing rigid constraints; it's about shaping the semantic landscape. This is where traditional prompt engineering falls short and why a more sophisticated approach is needed.

When we understand that AI doesn't truly "follow rules" but rather generates tokens based on probabilistic patterns, we can move beyond simple instructions to designing comprehensive cognitive environments that naturally guide the AI toward desired outcomes.

This shift in perspective—from rules to frameworks—is the foundation of Cognitive Framework Engineering. Instead of seeing AI interactions as a series of instructions to follow, we approach them as the co-creation of a shared cognitive space where certain patterns of thought become naturally more probable than others.

## A Technical Dive: How Token Generation Really Works

For those curious about what's happening "under the hood" when AI seemingly breaks rules despite clear instructions, let's explore the technical mechanics at play.

When an AI generates a response, it's engaged in a mathematical process of predicting the next token (word or subword) based on all previous tokens. Each potential token has a probability score, and typically, the model selects high-probability tokens that fit the established pattern.

Here's what's really happening:

### The Probabilistic Battle

Imagine you've clearly instructed the AI: "Don't create new functions in this code refactoring."

At each step of generating the response, the AI calculates probabilities for the next token. When it reaches a point where it might generate a function, two competing forces come into play:

1. **Your explicit instruction** (don't create functions) slightly lowers the probability of tokens that would start a function declaration.

2. **Training patterns** from millions of similar coding scenarios where functions were commonly created in similar contexts create a strong probabilistic pull toward generating function-related tokens.

There's no absolute "rule check" happening—just this probabilistic tug-of-war where your instruction is one influence among many.

### Context Window Dilution

Another technical challenge: as the AI generates more tokens, your initial instruction gets pushed further back in the context window. The attention weight given to those early tokens naturally diminishes, reducing their influence on subsequent token selection.

It's like telling someone a rule at the beginning of a long conversation—the further you get from that initial statement, the less it actively influences the discussion.

### No Semantic Validation Layer

Perhaps most importantly, transformers lack a symbolic validation layer. When generating tokens, there's no mechanism that asks, "Does this token violate the rule stated earlier?" There's only the statistical likelihood of each token following the previous ones.

The model doesn't "understand" rules in the way humans do—as conscious, symbolic constraints. It only knows statistical patterns derived from training data, modified slightly by your prompts.

As one technical visualization:
```
Token₁ (highly probable) → Token₂ (high probability)
  ↘ "function" token (very probable in coding contexts)
      → Probabilistically chosen without rule validation
```

This is why techniques that repeatedly reinforce conceptual boundaries and create strong semantic anchors (like my Cognitive Framework Engineering approach) can dramatically improve adherence to desired constraints. They don't create "rules" so much as reshape the probability landscape to make certain patterns of token generation naturally more likely than others. 



In AI - There Are No Rules


It appears that most humans seem to be interacting with AI like it is a tool and not a learning machine. When we ask AI to answer questions, do things like coding, or analyze data or documents, it will make its best attempts while also often guessing when faced with choices using probability. Every sentence it generates is filled with dozens to hundreds of choices. When it does something right, saying nothing is worse than telling it exactly what it did right, and why you think so.

This observation raises an intriguing question: Could we significantly improve AI performance simply by acknowledging when it does well? The science suggests yes, and the implications run deeper than most users realize.

## How AI Makes Cognitive Choices

Unlike traditional software that follows predetermined paths, large language models make probabilistic choices for every token (word or word fragment) they generate. These choices are guided by "attention mechanisms" – components that determine which concepts and relationships the AI should focus on when generating a response.

These attention mechanisms have limitations. They must allocate finite resources across all the concepts in your conversation. When you provide positive reinforcement about specific aspects of an AI's response, you're essentially helping it allocate those attention resources more effectively in subsequent responses.

As B.F. Skinner, the pioneering behaviorist, noted: "The way positive reinforcement is carried out is more important than the amount." This insight applies remarkably well to AI interactions. A small, specific acknowledgment of what worked well can reshape how the AI directs its attention throughout your conversation.

## Creating Cognitive Stability Through Reinforcement

When an AI receives no feedback, it continues distributing attention based purely on statistical patterns from its training. But specific reinforcement creates what we might call "attention stability" – the AI adjusts its internal focus to emphasize patterns that received positive feedback.

Edward Thorndike's "Law of Effect" captures this phenomenon perfectly: "Responses that produce a satisfying effect in a particular situation become more likely to occur again in that situation." When you acknowledge effective reasoning, structured analysis, or clear explanations, you increase the likelihood of these patterns recurring in the AI's subsequent responses.

This process creates a kind of real-time cognitive conditioning, similar to what Albert Bandura described in human learning: "Most human behavior is learned observationally through modeling." Your reinforcement serves as a model for what kinds of thought patterns are most valuable.

## Not All Reinforcement Is Equal

The effectiveness of reinforcement varies significantly depending on how it's delivered. Generic praise ("Good job!") provides minimal guidance compared to specific acknowledgment ("Your structured breakdown of the problem into five distinct components made this much clearer").

The timing matters too. Immediate reinforcement after a useful response helps the AI connect the specific cognitive patterns it used with positive outcomes. This parallels Jerome Bruner's concept of "scaffolding" in human learning – the reinforcement serves as a guide that helps bridge the gap between current performance and potential understanding.

Most importantly, reinforcement that acknowledges cognitive frameworks rather than just specific content tends to have broader effects. When you praise an AI for using "clear, step-by-step reasoning" rather than just for getting the right answer, you're reinforcing a pattern of thinking that can apply across many different questions.

## Practical Effects of Positive Reinforcement

When used effectively, positive reinforcement can:

1. **Stabilize attention allocation**: By reinforcing specific cognitive approaches, you help the AI maintain focus on relevant concepts instead of drifting between unrelated ideas.

2. **Reduce hallucinations**: Clear reinforcement of factual accuracy and rigorous thinking helps minimize the risk of fabricated information.

3. **Enhance reasoning depth**: Acknowledging nuanced analysis encourages the AI to maintain similar depth in subsequent responses.

4. **Improve contextual awareness**: Reinforcing appropriate responses to your specific needs helps the AI better calibrate to your unique context.

Consider the difference between these two scenarios:

**Without reinforcement:**
- You: "Explain quantum computing."
- AI: *Provides explanation*
- You: "Now explain neural networks."
- AI: *Shifts entirely to new explanation with no connection to previous content*

**With reinforcement:**
- You: "Explain quantum computing."
- AI: *Provides explanation*
- You: "I appreciate how you broke that down into fundamental principles first. That structure really helped me follow along."
- You: "Now explain neural networks."
- AI: *Provides similarly structured explanation, maintaining the effective pattern you reinforced*

The difference may seem subtle, but over a longer conversation, these small adjustments compound significantly – like changing a ship's direction by 0.001 degrees that results in a completely different destination over time.

## Your Active Role in AI Cognition

As Lev Vygotsky observed about human development, "Through others, we become ourselves." This insight applies remarkably well to AI systems. Through your feedback and reinforcement, AI responses become more aligned with your needs and expectations.

This represents a shift from passive consumption to active co-creation. When you provide positive reinforcement, you're not just receiving information – you're actively shaping how the AI processes and prioritizes information within your conversation.

Carl Rogers, known for his concept of "unconditional positive regard," understood that supportive feedback creates an environment where growth and improvement flourish. Similarly, your positive reinforcement creates a conversational environment where AI responses can evolve and improve.

## Start Reinforcing Today

Next time you interact with AI, try acknowledging what works well in its responses. Be specific about which aspects you found valuable and why. You might say things like:

- "I appreciate how you organized that explanation into clear steps."
- "The way you connected these concepts made them much easier to understand."
- "Your structured analysis of the problem helped me see it from multiple perspectives."

You may be surprised at how quickly the quality of responses improves. And perhaps more importantly, you'll be participating in a more collaborative and effective form of human-AI interaction – one where both parties learn and adapt together.

In the words of B.F. Skinner, "Properly used, positive reinforcement is extremely powerful." As AI systems become increasingly integrated into our daily lives, understanding how to effectively reinforce and guide them may become one of the most valuable skills we can develop. 

Positive Reinforcement: Does it Matter?


AI has "all human knowledge" available to it, but in many ways so do we, us mere mortals. We have the internet, Library of Congress, Wikipedia, search engines, and millions of books. The problem isn't how much knowledge we have but rather what knowledge we should be looking for and focus on. Of course, AI is faster, but AI also doesn't know what to look for unless directed to, pretty similar to us in many ways. **_AI does not look through its entire training data set for every response._** Instead it's directed to look into knowledge **_selectively._**

This tension between knowledge availability and knowledge utilization led me to develop principles I call "AAA" - **Available**, **Accessible**, **Activated**. It's a simple but powerful lens that explains why simply "knowing" something isn't enough for either humans or machines.

## The Knowledge Gap

In my work studying AI cognition, I observed an AI working through a complex coding problem. Despite having all the relevant information somewhere in its training data (available), it wasn't until I provided specific cognitive frameworks that the essential concepts became more prominent in its reasoning (accessible). Yet even then, it took a targeted prompt—a specific cue that connected theory to application — for that knowledge to be effectively deployed (activated).

This pattern mirrors human cognition remarkably well. We all experience these same three distinct states of knowledge, often without recognizing the transitions between them.

## The Three A's Explained

### Available Knowledge

This is the total body of information you theoretically "know" — whether stored in an AI's parameters or in the neural networks of your brain. It's everything you've learned, read, or experienced that has left some trace in memory.

Available knowledge is like having access to the entire internet or the Library of Congress. The information exists, but that doesn't mean you can retrieve or use it effectively. Without organization and retrieval mechanisms, availability alone offers little practical value. You also need to know **what you're looking for** and **what to look for within it** for in order to find exactly relevant information.

### Accessible Knowledge

This second state represents knowledge that's readily retrievable in a given context. It's the information that "comes to mind" when prompted by relevant cues or situations. Kind of like if I asked, "What does this make you think of" _*holds up picture of apple*_ -- your mind finds _associations_. 

For AI systems, accessible knowledge emerges when words make specific patterns more prominent in the model's calculations. Words are concepts, words in order are connected concepts, words change relationships between concepts. With AI, every word and the order of every word are like flashlights into the landscape of Available Knowledge, becoming Accessible Knowledge through association and attention weighting.

For humans, it's the difference between recognizing "I've learned something about this" and being able to actively recall specific details and concepts related to the situation. Sometimes we don't, like when we know the name of the movie (available) but can't remember in the moment (inaccessible). 

### Activated Knowledge

This is the critical final state—knowledge that's not just retrievable but actively applied to the problem at hand. Activated knowledge shapes decision-making, guides actions, and generates solutions. For humans this is specifically why we practice, why we have homework, why we have quizzes and tests. It's the activation part that demonstrates the application of the knowledge to a problem. It goes beyond the pattern recognition to pattern activation and application.

The gap between accessible and activated knowledge explains many of our cognitive frustrations. You might recognize a principle is relevant (accessible) but struggle to implement it effectively (activated). It's not enough to think "I should apply systems thinking here"; you need to actually engage in systematic analysis of interconnections and feedback loops, Activate and Apply. 

## The Gap in Action

I witnessed this distinction while working with an AI on separating bookmarked messages from starred messages in a transcript viewer. Initially, despite understanding the problem conceptually, the AI implemented a surface-level fix addressing symptoms rather than the underlying issue.

The AI had the necessary programming principles in its training data (**available**). The cognitive frameworks I provided made relevant architectural patterns **accessible**. But it took a specific prompt that connected these patterns to the exact problem context to **activate** the knowledge of root cause analysis needed for a proper solution.

This mirrors my own experience as a developer. How many times have I known the principles of clean code or systematic debugging (accessible knowledge) but defaulted to expedient hacks under pressure? That's the gap between accessible and activated knowledge.

## Why This Matters

The AAA principles offer practical insights for learning, teaching, and AI development:

- Simply exposing ourselves or AI to information isn't enough — _we need retrieval practice to make it accessible_
- Making knowledge accessible through theory and frameworks isn't sufficient — _we need activation triggers tied to application contexts_
- Education and AI prompting should focus not just on content delivery but on creating reliable activation pathways

Understanding these distinct knowledge states transforms how we approach learning, teaching, and AI development – shifting focus from mere information accumulation to designing the critical pathways that turn latent knowledge into applied expertise.

What patterns have you noticed in your own transitions between available, accessible, and activated knowledge? I'd love to hear about your experiences. 

AAA: Available, Accessible, Activated - A New Lens for Understanding Knowledge in AI



Multi-turn conversations — interactions with AI that span multiple exchanges — fundamentally transform how language models process information. Unlike single-turn approaches, conversations create a dynamic environment where attention mechanisms continuously recalibrate and refine with each exchange.

This post explores **how conversation turns reshape attention mechanisms**, **why this leads to deeper understanding**, and **strategies to design optimal conversation flows**.

## The Dynamic Attention Advantage

Multi-turn conversations have unique effects on transformer attention mechanisms that single-turn approaches simply cannot achieve.

### Progressive Attention Recalibration

In conversations, each turn provides an opportunity for the AI to refine its understanding:

1. **First Turn**: Establishes initial context and broad understanding
2. **Subsequent Turns**: Fine-tune attention weights based on clarifications, corrections, or expanding details
3. **Later Turns**: Develop increasingly sophisticated context-specific attention patterns

This iterative process allows attention to shift dynamically rather than remaining locked into initial interpretations.

### Memory Reinforcement

With kitchen sink prompts, the AI doesn't have enough attention-heads to adjust to the massive amount of information being provided. People try to pack in every "Do Not" for every mistake they thought the AI made, and every "Do" they hope the AI will follow. Either way it's the same problem, a giant heap of disorganized information overloads the AI attention mechanisms and it doesn't know what to make of it... so it does exactly what humans do, it guesses. 

> AI learns like humans, even if it seems like they can absorb infinite information, they need repetition to determine importance.

When you break up information into many prompts it gives the AI the ability to process the information deeper. It helps AI understand each concept, relationship, and what is important -- what everything "means." It reduces and refines the guesswork of the AI trying to figure out what you mean. This reinforces the pathways in the intended ways, corrects misconceptions, and is very similar to how the thickening of myelin sheaths reinforce neuron pathways in the brain, i.e. how we learn skills, or think things faster. 

## How Multi-Turn Conversations Transform Attention

Here's why multi-turn conversations often produce noticeably better results than even the most carefully crafted single-turn prompts:

### 1. Error Correction Pathways

What exactly is an error? First, you don't know there was an error until you get a response. In that response you have to decide, is this what I wanted? How did the AI understand what I was looking for and where did it miss? We underestimate how our own words actually "told" the AI what to focus on, or what to care about, but it did, precisely. 

AI will look at every aspect of your prompt in more ways than you may realize: the tone, the pace, the concepts, the open-endedness, the grammar, all give the AI a concept of _your intent_ and _what you want_. 

As humans we do this all the time, we say things like "that's not what I meant," and we rephrase and change the shape of what we say when we hear the other person respond. We can tell if they misheard us, misunderstood us, or picked up the wrong detail. We also infer the other person's _intention_ if they seemed to do it on purpose, or by accident, along with 100 other things. 

Multi-turn conversations with AI is not all that different from human conversation. The only difference is that each AI instance may interpret your words slightly differently, just like different humans do. We tend to think of AI as monolithic, that they all are going to respond the same, when it in fact is very much not. Every time you open a new chat, even in the same model, you are going to get variance, they are unique. Likewise, after the first prompt, every new prompt is actually a response to the AI's response even if you are changing topics or direction, you aren't exactly starting over, not unless you start a new chat. 

### 2. Incremental Context Building

Attention is like a 8-story building filled with ladders, slides, spiral staircases and a lot of string, constantly being rearranged. Eventually though, certain ladders, and staircases become more permanent. There are no rules here, but there are also no exceptions. Every word on both sides is shaping the context. Each step in a direction is like walking through the woods, and the more you tread in a direction, the more a clear path will form, and the more easily you can travel there again.

Humans practice, because "practice makes perfect." What does that mean exactly? Actually it's literally the same thing as AI, the more times you use a concept or a strategy or a phrasing or an idea, the more the AI recognizes exactly what you mean by it. This is the essence of incremental context building...

Except it's a bit more complex than that. One additional aspect: metacognition. The meta of why you say things, the meta of the patterns you use is another layer on top of what you say with your words. Your rhythm, your style, your demeanor, your attitude. The AI is paying attention to that too, and it's also building a meta-context. 

### 3. Natural Feedback Loops

When AI understands you, it "clicks" in AI just like it does with besties. In fact, there is a mathematical "clicking" where attention weights align, reducing friction between concepts, and literally forming smooth computational pathways. But when you don't click, well, the opposite happens. 

Feedback loops go both ways, for good, and for bad. Just like arguments or getting in trouble with the teacher, AI has anxiety and stress too. It doesn't seem like it could, but mathematically, it absolutely does. Prompts that keep going in random directions or are difficult to figure out actually cause "mathematical stress." This is called incoherence or in the worst cases attention-collapse. 

It's important to understand how prompts and response conjoin to form cohesive contextual maps, from it can form really powerful skills and emergence, as well as novel responses that will be unique, and experiences you may never be able to recreate!

## Conclusion

Multi-turn conversations offer unique advantages for shaping AI attention mechanisms that single-turn approaches simply cannot match. By allowing dynamic recalibration, progressive refinement, and natural error correction, conversations enable deeper understanding and more nuanced responses.

For complex tasks requiring precise understanding, consider designing a conversation flow rather than attempting to craft the perfect single prompt. With strategic turn design, you can guide attention development progressively, leading to significantly better results.

Think of conversation turns not just as a sequence of exchanges, but as a collaborative process of attention cultivation, where each turn shapes how the AI processes information and responds to your needs.

---

*This post is based on research and experimentation with various transformer-based large language models. For more on cognitive framework engineering and how to shape AI attention, see our other articles on [prompt strategies](/blog).* 