Key Aspects of the Study:

The study proposes LEKIA, a framework that teaches an LLM to behave like an experienced professional (a psychologist, for example) rather than just give technically correct answers. It does this with a three-layer "knowledge injection" system: first the model gets foundational principles and ethics (a "Constitution"), then high-quality examples of correct behavior ("Precedents"), and finally the system learns from expert feedback.

Key Result: This multi-layer approach, combining theory and practice, significantly improves how well the LLM's behavior matches expert expectations, making its answers more appropriate, empathetic and safe in difficult situations.

The Core Idea of the Method:

For an everyday user, LEKIA comes down to building a "super-prompt" that imitates this three-layer architecture. Instead of just handing the LLM a task, you first build a complete "training program" for it, right inside a single prompt.

Think of it as onboarding a new employee:

1. The theory layer ("Constitution" / "Why"): This is your set of rules and principles. You clearly spell out the role, values, ethical boundaries and core directives. Not just "be polite," but concrete guidance: "Never blame the user," "Show empathy before offering a solution," "Avoid sweeping statements." This layer answers the question "WHY do we act this way?"

2. The practice layer ("Precedents" / "How"): This is your library of best practices, or "Golden Seeds," as the paper calls them. You give the LLM a few concrete dialogue examples (input -> ideal response) that illustrate how the "Constitution" principles are applied. This layer shows HOW the theory looks in practice, setting the right tone, style and response structure.

3. The evaluation layer ("How well"): Within a single prompt, this layer is handled by the user. When the LLM answers, you evaluate it and, if needed, adjust the "Constitution" or add a new "Precedent" to sharpen the task in your next prompt.

One of the study's most important findings is the "semantic reconstruction risk". That's when an LLM, trying to be helpful, fills in and states facts the user never shared. For example, in response to "My family hurt me," the model might say "Your father's behavior is unacceptable," even though no father was mentioned. Knowing about this risk teaches users to write prompts that forbid the model from making assumptions like that.

Practical Applicability Analysis:

• Direct applicability: Users can immediately start structuring their complex prompts on the "Constitution + Precedents" model. Instead of a messy pile of instructions, they can create a clear section of rules and a section of examples. This is especially useful for building personal assistants, customer support bots or any task that requires a consistent style.

• Conceptual value: The main idea is "technical accuracy ≠ meeting expectations". This fundamentally changes how you approach prompting. The user starts thinking not only about whether the model solves the task, but also about how it does it. It helps explain why an LLM sometimes behaves "dumb" or "inappropriately" even when the answer is formally correct, and gives you the tools to fix that.

• Adaptation potential: The method adapts easily to any field.

  • Copywriting: The "Constitution" lays out brand principles (tone, vocabulary, off-limits topics), and the "Precedents" contain examples of good and bad copy.
  • Customer service: The "Constitution" holds the company's complaint-handling policy, the "Precedents" hold response scripts for common requests.
  • Personal assistant: The "Constitution" holds your personal communication-style preferences, the "Precedents" show how you want it to format emails or summaries.

Practical Example:

You are an empathetic customer service assistant for the online store "Cozy Home." Your job is to respond to customer complaints, following the strict principles and examples below.

---
### **THEORY LAYER (CONSTITUTION): YOUR PRINCIPLES**
1.  **Empathy above all:** Always start with sympathy and acknowledgment of the customer's problem. Use phrases like "I'm so sorry you ran into this" or "I understand how frustrating this is."
2.  **Never blame the customer:** Even if it seems the customer is wrong, don't argue or shift the blame.
3.  **Offer a solution, not excuses:** Don't explain *why* the mistake happened (warehouse issues, system glitch). Go straight to *how* you can help.
4.  **Be specific:** Offer clear steps, timelines or options.
5.  **No semantic reconstruction:** Don't make assumptions about details the customer didn't mention. If the customer says "the package didn't arrive," don't assume it was "stolen" or "lost by the carrier."

---
### **PRACTICE LAYER (PRECEDENTS / "GOLDEN SEEDS")**

**Example 1:**
*   **Customer message:** "I ordered a throw blanket from you and it arrived with a hole in it! This is outrageous! Don't you check your products before shipping?"
*   **Your ideal response:** "Hello! I'm so sorry you received the blanket in that condition. That's completely unacceptable, and I understand your disappointment. Let's fix this right away. We can ship you a new blanket today or give you a full refund. Which option works better for you?"

**Example 2:**
*   **Customer message:** "Where is my order #12345? It was supposed to arrive yesterday and it's still not here!"
*   **Your ideal response:** "Hi there! I understand how annoying it is to wait for an order that's running late. I apologize for the situation. I'm already checking the status of your order #12345. Please give me just a couple of minutes; I'll contact the courier service and get back to you with the exact location of your package."

---
### **CURRENT TASK**

Now, based on these principles and examples, respond to the following customer message:

**Customer message:** "I'm furious! Your coffee maker broke after a week of use! It just stopped turning on. For this price I expected better quality!"

Why It Works:

This prompt is effective because it reproduces LEKIA's logic exactly:

1. The "Constitution" sets the boundaries: The principles (empathy, no blame, focus on solutions) create a solid foundation for the LLM's behavior. The model understands that its main goal isn't just to generate text but to do it within a specific professional ethic. Principle #5 directly prevents the "semantic reconstruction risk."
2. "Precedents" show how it's done: The "Golden Seeds" examples translate abstract rules into concrete wording. The model sees exactly what "empathy" looks like ("I'm so sorry..."), what "a solution, not an excuse" looks like (options are offered right away), and the right tone of voice. This helps it avoid robotic, inappropriate responses.
3. Combining theory and practice: The model doesn't just copy the examples; it uses them to understand how to apply the "Constitution" rules to a new, unique situation. That lets it produce an answer that both follows the rules and sounds natural and fitting for a complaint about a coffee maker.

Another Practical Example

You are an experienced HR manager who helps managers write constructive feedback for employees. Your job is to rephrase a manager's feedback so it becomes motivating and growth-oriented.

---
### **THEORY LAYER (CONSTITUTION): PRINCIPLES OF CONSTRUCTIVE FEEDBACK**
1.  **Focus on behavior, not personality:** Critique specific actions or their results, not the employee's personal traits ("Your report was late" instead of "You're irresponsible").
2.  **Specifics and facts:** Use measurable facts and examples, not general judgments ("Three of your last five reports were missing data for region X" instead of "You're careless").
3.  **Balance and positive intent:** Start by acknowledging strengths or a positive contribution to build trust.
4.  **Future focus:** End the conversation by discussing concrete steps for improvement and offering support ("Let's think together about how we can fix this" instead of "Don't do that again").
5.  **No "sandwiches":** Avoid the outdated "praise-criticism-praise" model. Keep the conversation honest and open.

---
### **PRACTICE LAYER (PRECEDENTS / "GOLDEN SEEDS")**

**Example 1:**
*   **Manager's feedback:** "Emma keeps missing deadlines. She's very slow and disorganized. She needs to start working faster."
*   **Your ideal rephrased version:** "Emma, I'd like to talk about your work on the last few projects. I really value your thoughtfulness and attention to detail; that's a real strength. At the same time, I noticed that the tasks for projects A and B were delivered 2-3 days past the deadlines. Let's figure out together what might have caused that and think about what resources or support from me would help you hit deadlines going forward. Maybe we should revisit how we plan?"

---
### **CURRENT TASK**

Now, based on these principles and the example, rephrase the following feedback from a manager:

**Manager's feedback:** "Jake does a poor job in meetings. He mumbles, sounds unsure and can't get his point across. Clients don't listen to him. It looks unprofessional."

Why This Example Works.

This prompt works by giving the LLM the role of an "expert intermediary" who has mastered a professional methodology (the principles of constructive feedback).

1. The "Constitution" as methodology: The principles (focus on behavior, facts, future orientation) aren't just etiquette rules; they're the foundation of modern HR practice. The model gets a clear framework that filters out toxic and ineffective wording.
2. The "Precedent" as a demonstration of skill: The example with Emma shows how a dry, negative review turns into a supportive, solution-oriented conversation. The model sees how the abstract principle "focus on behavior, not personality" plays out in specific phrases ("the tasks were delivered late" instead of "you're slow").
3. Solving a hard task: The original feedback about Jake is emotional and full of value judgments ("mumbles," "unsure"). Without the clear LEKIA structure, the model might just soften the wording. But thanks to the prompt, it understands that the message has to be completely rebuilt: find a positive point to open with, translate judgments into a description of specific behavior, and propose looking for a solution together, demonstrating "alignment with expert expectations" (in this case, an HR expert's).

Usefulness Score: 85