Key Aspects of the Study:

The researchers compared four different ways of framing a task (a prompt) for ChatGPT when classifying texts into predefined categories. They found that a prompt forcing the model to first analyze the text, then explain its reasoning, and only then give a final answer (the "Step-by-Step Task Decomposition" method) works significantly better and more reliably than plain instructions or instructions with examples.

Key Result: Making an LLM reason step by step before it answers can substantially raise the accuracy and reliability of text classification.

The Core of the Method, Explained:

The idea behind "Step-by-Step Task Decomposition" is simple: don't let the LLM answer right away. Instead, make it walk the same mental path a human expert would. Rather than just asking "Which category does this text belong to?", you hand the model a clear algorithm to follow.

That algorithm usually includes: 1. Analyzing the input: First, the model identifies and pulls out the key facts, entities, themes, or arguments from the text. 2. Matching against criteria: Next, the model takes those extracted facts and matches them against the definitions of the categories you provided. 3. Spelling out the reasoning: At this stage, the model explicitly writes down its chain of logic: "Since the text mentions X, which fits the definition of category Y, I believe this category applies." 4. Delivering the final answer: Only after all the previous steps does the model give its final, structured answer.

This approach swaps the model's "gut-feeling" guesswork for a structured, transparent, and verifiable process. It forces the LLM to "think out loud," which sharply cuts the error rate, especially on tasks with ambiguous or overlapping categories.

Practical Applicability:

• Direct applicability: Anyone can use this approach right away for any task that involves analysis and classification: sorting support emails, analyzing customer reviews, categorizing news, or screening candidates by their resumes. Just build a "Step 1, Step 2, Step 3, Result" structure into your prompt.

• Conceptual value: The study hands you three powerful ideas:

  1. Process beats output: The quality of an LLM's final answer depends directly on the quality of the reasoning process you give it.
  2. Reasoning > Examples: For complex analytical tasks, making the model reason through an algorithm works better than just showing it a few examples (few-shot).
  3. Context "fatigue" (Instruction Decay): Within a single session, an LLM can start performing worse after 30-40 repetitive operations. So for bulk jobs, either split them across several sessions or "refresh" the instruction periodically.

• Adaptation potential: The mechanism adapts easily. Instead of "classification," it can be "evaluation," "comparison," or "risk spotting." For example, when reviewing a contract you could use these steps: "1. Find the clauses about penalties. 2. Find the clauses about deadlines. 3. Assess the risk in each clause. 4. Write a final conclusion." The principle of decomposition and reflection stays the same.

A Practical Example:

You are an experienced customer success manager. Your task is to analyze a customer review of our mobile app and classify it by sentiment and main topic.

**Classification categories:**
*   **Sentiment:** Positive, Neutral, Negative
*   **Main topic:** Payment issue, Improvement suggestion, App error/bug, Praise for the interface

**Review text to analyze:**
"Overall the app looks great and it's nice to use. But for the second time now I can't pay for my subscription with a card, it throws some kind of error. I have to pay through the website instead, which is a hassle. Please add Apple Pay, that would be awesome!"

**Complete the task strictly step by step:**

**Step 1: Sentiment analysis.**
Analyze the text and determine the overall sentiment. Find phrases in the text that support your choice.

**Step 2: Main topic analysis.**
Analyze what the customer is focused on. Which problem or idea is central to their message? Provide quotes.

**Step 3: Reasoning.**
Briefly explain why you chose these specific categories based on your analysis in steps 1 and 2.

**Step 4: Final result.**
Provide the final answer in JSON format.

{
  "tonality": "...",
  "main_topic": "..."
}

Why This Works:

This prompt is effective thanks to several mechanics described in the study:

1. Task decomposition: Instead of one complex request like "Analyze and classify," the task is broken into four clear, sequential steps. This steers the model's "attention" and keeps it from drifting.
2. Forced reflection and evidence extraction: The instructions "Find phrases that support your choice" and "Provide quotes" make the model ground its conclusion in specific data from the text instead of guessing. This mirrors the study's "link relevant textual evidence" requirement.
3. Transparent logic: Step 3 ("Reasoning") makes the model state its chain of logic explicitly, which raises the odds of a correct conclusion and lets a human easily check the LLM's logic.
4. Structured output: Requiring a JSON result (Step 4) guarantees predictable data that's easy to process downstream, separating the "reasoning" from the "answer."

Another Practical Example

You are an experienced HR specialist. Your task is to do an initial screening of a candidate's application for a "Content Manager" role and give a short conclusion.

**Key job requirements:**
1.  2+ years of experience with CMS platforms (WordPress, Squarespace).
2.  A portfolio with writing samples (articles, posts).
3.  Basic SEO knowledge.

**Candidate's cover letter:**
"Hi there! I'm interested in your content manager opening. I've been working in this field for three years now, mostly with WordPress. I've handled filling websites with content, written blog articles on a range of topics, and I'm somewhat familiar with keyword research. You can see examples of my work here: [portfolio link]."

**Run the analysis strictly step by step:**

**Step 1: Check requirement #1 (CMS experience).**
Find information about CMS experience in the text. Assess whether it meets the requirement (2+ years).

**Step 2: Check requirement #2 (Portfolio).**
Find any mention of a portfolio in the text. State whether the candidate provided a link.

**Step 3: Check requirement #3 (SEO knowledge).**
Find any mention of SEO knowledge in the text. Assess whether it meets a basic level.

**Step 4: Final conclusion.**
Based on the analysis in steps 1-3, give a short conclusion: "Recommend for interview" or "Reject." Briefly justify your decision.

Why This Example Works.

This example shows how the method adapts to evaluation and decision-making tasks, not just classification.

1. Decomposition by criteria: The task "evaluate the candidate" is broken into a series of smaller, concrete checks against each key job requirement. That turns a fuzzy evaluation into a concrete checklist.
2. Lower cognitive load: The model doesn't have to juggle all three requirements "in its head" at once. At each step it focuses on just one aspect, which improves accuracy and reduces the risk of missing an important detail.
3. A justified decision: The final conclusion in Step 4 isn't pulled out of thin air. It follows logically from the previous analysis steps. The "Briefly justify" requirement makes the model confirm its logic one more time, which lines up directly with the study's principles and makes the final verdict more reliable.

Usefulness Score: 95