Built by TARAhut AI Labs
Generative AI & Prompt Engineering · Session 06 · tarahutailabs.com
© 2026 TARAhut AI Labs
4 words that change everything: "Think step by step." Chain-of-thought prompting teaches AI to REASON, not just respond. Watch wrong answers become right answers. "Sochde jao, kadam kadam — AI nu sochna sikhao."
"If a shirt costs ₹800 after a 20% discount, what was the original price?" GPT-4o gets this wrong without CoT. Add "Think step by step" and it gets it right. Those 4 words change everything. "4 shabd naal galat jawab sahi ho jaanda hai."
"A shirt costs ₹800 after 20% discount. Original price?"
AI might answer: "₹960" (20% of 800 = 160, so 800+160=960).
✕ Wrong! If original is ₹960, 20% off = ₹768, not ₹800.
"Think step by step. A shirt costs ₹800 after 20% discount."
Step 1: 800 = original x (1 - 0.20)
Step 2: 800 = original x 0.80
Step 3: original = 800 / 0.80 = ₹1,000
Normally, AI jumps from question to answer in one leap. CoT forces it to show its work, step by step. Each step checks the previous one. Like showing your work in math class. "Jive math class vich kaam dikhana — har step check hunda hai."
Next: The 3 levels of chain-of-thought prompting.
From "just add 4 words" to providing complete reasoning structures. Each level gives you more control. "3 levels — har level zyada control dinda hai."
Just add Think step by step or Let's work through this systematically. No examples needed. Easiest to use, surprisingly effective. Works for 70% of reasoning tasks. "Sirf 4 shabd jod do."
You provide the reasoning structure: First, identify X. Then, calculate Y. Then, compare Z. Finally, conclude. You define the steps, AI follows them. More reliable for complex tasks. "Tusi steps define karo, AI follow kare."
You provide a complete worked example WITH reasoning, then ask AI to follow the same pattern for a new problem. Most reliable, highest quality. "Pehle example dikhaao, phir nava problem do."
Use CoT for: Math, logic, multi-step analysis, decision-making, code debugging, legal reasoning, financial projections.
Skip CoT for: Simple facts ("What is the capital of France?"), translation, summarization, creative writing. "Jado sochna zaruri hai, CoT use karo. Jado simple kaam hai, zarurat nahi."
Next: Tree-of-Thought — explore MULTIPLE reasoning paths.
CoT is linear: one path of reasoning. Tree-of-Thought (ToT) explores MULTIPLE paths simultaneously, then picks the best one. Like a chess player thinking 3 moves ahead on different strategies. "CoT = ik raasta. ToT = 3 raaste explore karo, sabton vadiya chuno."
Next: Apply CoT/ToT to YOUR domain problems.
Apply chain-of-thought to YOUR complex tasks. Find 3 problems where AI gives wrong/weak answers and fix them with CoT.
8 questions from a pool of 16.
✅ "Think step by step" fixes most reasoning errors
✅ 3 CoT levels: zero-shot (easy), manual (structured), structured (example-based)
✅ Tree-of-Thought explores multiple strategies before choosing
✅ CoT + CRISP = professional-grade prompt engineering
✅ Use CoT for reasoning, skip for simple tasks
"Kal assi AI nu examples naal sikhaange. Few-shot prompting naal classification, extraction, te formatting tasks dramatically better hunde ne. 3-5 examples deo, AI pattern follow kare."