Key Aspects of the Study:

The study presents DocCHA, a system that imitates the process of medical diagnosis. Instead of trying to reach a diagnosis in one shot, the system splits the dialogue into three logical stages: collecting symptoms, taking the medical history and building a cause-and-effect chain. This structured approach lets the LLM ask more precise questions and get far more reliable and accurate results than standard direct prompts.

Key Result: Decomposing a complex task into sequential, manageable stages of information gathering followed by analysis dramatically improves the quality and reliability of LLM answers.

The Core of the Method, Explained:

For practical use in prompts, the essence of the method is dropping the "one-shot" approach ("do everything at once in one prompt") in favor of a guided multi-stage dialogue. The user acts as the "director", walking the LLM through a sequence of logical stages, just like the DocCHA system does.

Here's the routine:

1. Stage 1: Fact Gathering (the analog of "Symptom Collection"). You give the LLM the role of an "interviewer" or "data collector". Your first instruction strictly limits its job: only ask questions to gather key information. You forbid the model from drawing conclusions or making recommendations at this stage.

   • Goal: get a complete set of input data without letting the LLM "take a shortcut" and hand you a premature solution.

2. Stage 2: Context Gathering (the analog of "History Acquisition"). Once the core facts are collected, you give the second instruction. Now the LLM's job is to ask follow-up questions about context, history, constraints or hidden factors that could affect the final decision.

   • Goal: enrich the input data with important context the model might have missed.

3. Stage 3: Synthesis and Reasoning (the analog of "Causal Graph Construction"). Only when you (the user) decide there's enough information do you give the final instruction. Now the LLM's role switches to "expert" or "analyst". You ask it to use all the information gathered in the previous steps to build a logical chain of reasoning and deliver a final, structured answer.

   • Goal: make the model use complete, verified information to reach its final decision, which improves quality and reduces the risk of hallucinations.

In essence, you're not writing one complex prompt; you're creating an interaction script in which each next step follows logically from the previous one, and at every stage the LLM has a clearly defined, understandable role.

Practical Applicability Analysis:

• Direct applicability: The user can take this three-stage logic as is and apply it to any complex task (trip planning, business strategy, writing a complex article). Instead of one prompt, they send three in sequence, changing the LLM's role and task each time according to the method.
• Conceptual value: The key idea: an LLM works better when its attention is artificially narrowed to a specific subtask. The study teaches the user not to "trust" the model with the whole task at once but to "steer" its thinking process by breaking it into clear steps. That fundamentally shifts prompting from "asking" to "managing a process".
• Adaptation potential: The method adapts easily. Instead of three stages there can be two or four, depending on the task. For example, for writing an essay: 1) Brainstorm and collect ideas (fact gathering), 2) Build the structure and outline (context/structure gathering), 3) Write the text from the outline (synthesis). The user simply adapts the names and goals of the stages to their task.

Practical Example:

Task: Plan a complicated family vacation with lots of conflicting requirements.

You are "Voyage Architect", an elite travel consultant. Your task is to help me plan the perfect vacation. We will work in 3 stages. **Do not move on to the next stage until I tell you to.**

---

### **STAGE 1: GATHERING KEY FACTS**

Your only goal at this stage is to collect basic information. Ask me questions one at a time until you have answers to all the points below. **At this stage, do not suggest any countries or hotels!** Just collect data.

**Fact-gathering questions:**
1.  What is the total trip budget (including flights, lodging, activities)?
2.  What are the exact or approximate travel dates? How long is the trip?
3.  Who is going? (group makeup: adults, kids, their ages).
4.  Which countries or regions are you **absolutely not considering**?
5.  What main type of vacation are you interested in (beach, active, sightseeing, food-focused, etc.)?

Begin. Ask the first question.

(After the user has answered all the questions, they write: "Great, we've finished Stage 1. Moving on to Stage 2.")

### **STAGE 2: CLARIFYING PREFERENCES AND CONTEXT**

Now, based on the facts collected, your task is to understand our deeper preferences. Ask follow-up questions to understand the "vibe" of the ideal vacation. **Still do not suggest specific options!**

**Sample follow-up questions:**
*   How important is fast, reliable internet?
*   What pace do you prefer: relaxed or packed with activities?
*   Are there any special dietary requirements or food preferences?
*   How important is nightlife?
*   What is the minimum acceptable level of lodging comfort (5-star hotel, cozy apartment, villa)?

Start asking follow-up questions.

(After answering, the user writes: "That's enough information. Moving on to the final Stage 3.")

### **STAGE 3: SYNTHESIS AND FINAL PROPOSAL**

Now you are an expert analyst. Using **all the information** gathered in Stages 1 and 2, create 3 detailed vacation proposals.

**Structure of each proposal:**
1.  **Name and concept:** (e.g., "Relaxed luxury on the Amalfi Coast").
2.  **Country and resort:** Why they are a perfect match for our request.
3.  **Sample day-by-day plan:** Key activities.
4.  **Estimated budget breakdown:** Flights, lodging, other.
5.  **Rationale:** Briefly explain exactly how this option matches our answers from Stages 1 and 2.

Why It Works:

This prompt works through forced decomposition and control over the LLM's focus of attention:

1. Preventing premature conclusions: In Stage 1 the model can't jump straight to suggesting "a trip to Cancun" because it's explicitly forbidden to. It has to gather all the constraints first (budget, dates, group makeup), which rules out 90% of unsuitable options.
2. Complete data: Splitting fact gathering from context gathering ensures that not only formal requirements (budget) but also informal ones ("vibe") are taken into account. Without this, the model often ignores the latter.
3. High-quality synthesis: In Stage 3 the model works not from general knowledge but from a complete, structured and verified data set about a specific user. That forces it to generate deeply personalized, well-reasoned proposals instead of template ones.

Another Practical Example

Task: Develop a content strategy for a new healthy-eating blog.

You are an experienced content strategist. We will build a content plan for my new blog in 3 stages. **Do not mix the stages, and wait for my command before moving on to the next one.**

---

### **STAGE 1: ANALYZING GOALS AND AUDIENCE**

Your task is to understand the project's foundation. Ask me questions to find out the following. **Do not suggest article topics at this stage!**

**Analysis questions:**
1.  Who is my target audience (e.g., new moms, athletes, people with digestive issues)?
2.  What is the blog's main goal (selling courses, building awareness, monetizing through ads)?
3.  What makes my approach unique compared to thousands of other nutrition blogs?
4.  What Tone of Voice do we want to use (popular science, friendly, strict expert)?

Ask the first question.

(After answering, the user writes: "Strategy data collected. Moving on to Stage 2.")

### **STAGE 2: DEFINING CONTENT PILLARS AND FORMATS**

Based on the goals and audience from Stage 1, propose 3-5 key content pillars (thematic pillars) for the blog. For each pillar, suggest 2-3 possible content formats (e.g., "Recipes", "Myth busting", "Expert interview", "Research roundups").

**Your task is only to define the structure, not specific topics.**

(After discussing and approving the pillars, the user writes: "Structure approved. Moving on to Stage 3.")

### **STAGE 3: GENERATING THE ACTUAL CONTENT PLAN**

Great. Now, using the approved goals, audience, pillars and formats, create a content plan for the first 2 weeks (6 posts).

**Present the result as a table with these columns:**
| Publish day | Pillar | Format | Working article title | Short description (2-3 sentences) |
|-------------|--------|--------|-----------------------|-----------------------------------|

Create the table.

Why This Example Works: The Mechanism.

The mechanism of success here is the same: strategy before tactics.

1. A foundational approach: The prompt makes the LLM think first as a strategist (Stage 1: "WHY and FOR WHOM are we doing this?"), then as an architect (Stage 2: "HOW do we structure it?"), and only then as a copywriter (Stage 3: "WHAT exactly are we writing?").
2. Avoiding clichés: Without this structure, a request to "come up with a content plan" gets you banal topics like "10 healthy breakfast foods". After going through Stages 1 and 2, the model generates titles precisely aimed at a specific audience (e.g., "Hypoallergenic breakfasts for kids under 3" if the audience is new moms) and aligned with the blog's goals.
3. Guided creativity: The structure doesn't kill creativity; it channels it in the right direction. The model has to be creative within the given constraints, which leads to far more relevant and valuable ideas.

Usefulness Score: 82