SESSION 2 · ~40M
Chain of thought & reasoning
Asking a model to reason before answering is the single highest-leverage prompting technique known — and the reason is mechanical, not mystical.
Models reason better when they generate intermediate steps before the final answer. This is chain-of-thought, and it works because each token the model writes becomes part of its own context — it effectively gets to 'think out loud' before committing.
Why steps beat leaps
A model predicting the answer in one shot has to be right about everything at once. A model that writes its reasoning first can break a hard problem into easier sub-predictions, each of which is more likely to be correct. The cumulative accuracy is dramatically higher — especially for math, logic, and multi-step planning.
Patterns that work
- 'Think step by step before answering.' — the classic trigger.
- 'First list the assumptions, then solve.' — surfaces hidden premises.
- 'Decompose: break this into sub-problems, solve each, then combine.' — for complex tasks.
- 'State your plan, then execute it.' — for agentic or multi-step work.
Decomposition is the general principle behind all of these: the smaller and more atomic each step, the more reliable the model becomes. Whenever a model fails a complex task, your first fix is to ask it to break the task into smaller explicit steps.
Caveat: chain-of-thought uses more tokens (cost) and the reasoning can drift if the task is genuinely outside the model's competence. Steps improve reliability within the model's reach; they don't grant capabilities it lacks.
TRY IT
Give a model a multi-step word problem. First ask for the answer directly. Then re-ask with 'reason step by step'. Compare. The gap you see is the entire case for chain-of-thought in one demonstration.
CHECK YOUR UNDERSTANDING
Mechanically, why does chain-of-thought improve model accuracy?
What is the general principle behind all effective chain-of-thought patterns?
What is the stated caveat about chain-of-thought?
NOTEBOOKLM
Connect your notebook
Paste a NotebookLM share link to embed your notebook for this module.