Quest 19 - Chain-of-Thought Prompter

Quest 19: Chain-of-Thought Prompter

medium 25 minutes

---

🎯 Learning Objectives

• ✅ Understand chain-of-thought prompting and why step-by-step reasoning improves accuracy
• ✅ Structure prompts that elicit explicit reasoning from LLMs
• ✅ Break complex problems into verifiable reasoning steps
• ✅ Apply CoT patterns to improve performance on math, logic, and multi-step tasks

---

📖 Concept: Chain-of-Thought Prompting

Chain-of-Thought (CoT) prompting คือ technique ที่สั่งให้ LLM “คิดให้เห็น” แทนที่จะข้ามไปคำตอบโดยตรง ผลลัพธ์คือ accuracy ที่สูงขึ้นมากสำหรับ tasks ที่ต้องใช้ reasoning

แทนที่จะถาม “2 + 2 เท่ากับเท่าไหร่?” ตรง ๆ CoT จะถามว่า “คิดทีละขั้น: 1) 识别 operation, 2) คำนวณ, 3) ตรวจสอบคำตอบ” ทำให้ model ไม่ข้าม step และมี chance แก้ไขตัวเอง

เหมือน martial arts: แทนที่จะจำ technique ทั้งหมดทีเดียว CoT คือการ breakdown เป็นท่าเล็ก ๆ ที่ฝึกทีละ step จนกลายเป็น muscle memory

---

⚙️ How It Works

The CoT Workflow

1. Receive problem and reasoning steps
         ↓
2. Format prompt with structured steps
   ├── Include the problem statement
   ├── List reasoning steps explicitly
   ├── Add "Let's think step by step" pattern
   └── Separate each step clearly
         ↓
3. Return formatted prompt
         ↓
4. LLM processes each step sequentially
         ↓
5. More accurate final answer

Why CoT Works

Approach	Process	Accuracy
Direct prompt	Problem → Answer	Low for complex tasks
CoT prompt	Problem → Step 1 → Step 2 → … → Answer	High

CoT forces the model to use its intermediate tokens for computation, not just storage.

---

💡 Example: Building a CoT Prompt

Step 1: Basic structure

function buildCoTPrompt(problem, steps) {
  let prompt = `Problem: ${problem}\n\n`;
  prompt += "Let's think step by step:\n\n";

  steps.forEach((step, i) => {
    prompt += `Step ${i + 1}: ${step}\n`;
  });

  prompt += "\nNow provide your answer based on the above reasoning.";
  return prompt;
}

Step 2: Usage example

const prompt = buildCoTPrompt(
  "What is 2+2?",
  ["Identify the operation (addition)", "Compute 2+2 = 4", "Verify: 4 is reasonable"]
);

Step 3: Verify the output

node test.js
# Test: contains the problem
# Test: contains step-by-step indicator
# Test: contains all steps
# Test: not just the problem passed through

---

⚠️ Common Mistakes

Mistake 1: Returning the problem as-is

return problem without any structuring → CoT requires explicit step structure. Pass-through defeats the purpose.

Mistake 2: Not including a reasoning indicator

Missing “Let’s think step by step” or similar phrase → This phrase triggers the model’s reasoning mode. Without it, CoT is less effective.

Mistake 3: Steps that are too vague

“Think about it” instead of specific reasoning steps → Each step should be a concrete, verifiable action.

Mistake 4: Not handling empty steps

Crashing when no steps are provided → Always handle the edge case gracefully — return at least the problem with the CoT trigger phrase.

---

📝 Knowledge Check

📝 Knowledge Check

Q1:What is chain-of-thought (CoT) prompting?

Asking the model to generate code step by stepStructuring prompts to elicit explicit step-by-step reasoningChaining multiple API calls togetherUsing multiple models in sequence

Q2:Why does CoT prompting improve accuracy on math problems?

It gives the model more tokens to compute withIt forces the model to reason through each step instead of jumping to an answerIt automatically verifies the answerIt uses a different model architecture

Q3:What should a CoT prompt include even when no steps are provided?

An error messageNothing — return empty stringAt least the problem statement and a 'step by step' trigger phraseA list of default generic steps

---

🏋️ Quest: Chain-of-Thought Prompter

Now it’s time to build structured reasoning prompts!

1. Download ไฟล์เริ่มต้นของ quest:

   npx bluebeltdojo download quest-19-chain-of-thought
   cd quest-19-chain-of-thought

2. เปิด problem.js ใน editor ของคุณพร้อมความช่วยเหลือของ AI

3. Implement buildCoTPrompt(problem, steps) — สร้าง structured reasoning prompt

4. ทดสอบ edge cases: empty steps, single step, multiple steps

5. ตรวจสอบ solution ของคุณ:

   node test.js

6. When all tests pass, submit your solution:

   npx bluebeltdojo submit

💡 Tip: CoT prompting คือ one of the simplest but most powerful techniques — ลองกับ math problems แล้วจะเห็นผลทันที

---

คำใบ้

• อ่าน instructions ใน problem.js อย่างละเอียด
• ตรวจสอบ: prompt ต้องมี “step by step” indicator และ step ทั้งหมด
• ถ้าติดขัด ลองอ่าน “Common Mistakes” อีกครั้ง — อย่าดู solution โดยตรง