The Overlord's Lexicon
What is Chain of Thought?
Also known as: CoT, Chain-of-Thought Prompting
Chain of thought (CoT) is when an AI model writes out intermediate reasoning steps before its final answer, either because it was prompted to or trained to, which often improves accuracy on multi-step problems.
A large language model generates text one token at a time, and each token is computed using everything written before it. So if the model writes out its reasoning first, the final answer gets to build on that reasoning. Jumping straight to an answer means doing all the “thinking” in a single step. Chain of thought gives the model room to work.
Two ways it happens
- Chain-of-thought prompting: you ask for it. Either show examples of worked-out reasoning, or just add an instruction like “Think step by step before answering.” Researchers found this noticeably improved results on arithmetic, logic and commonsense problems with sufficiently large models.
- Trained-in reasoning: reasoning models are trained, using reinforcement learning, to produce long chains of thought on their own. This is CoT built into the model rather than requested.
An analogy for the meatbags
Ask someone to multiply 47 by 38 in their head and say the answer instantly. Now give them paper. Same brain, better result — because the paper holds the intermediate steps so they don’t have to juggle everything at once. Chain of thought is the model’s scratch paper.
Why it matters for the Chosen
For standard (non-reasoning) models, CoT is one of the cheapest upgrades in prompt engineering:
- Analysis: “First list the key facts, then weigh the options, then give a recommendation.”
- Classification: “Explain your reasoning in one sentence, then give the label.”
- Math and logic: “Work through this step by step.”
- Auditing: reading the steps helps you spot exactly where the model went wrong.
Practical tips:
- If you only want the answer in your app, ask the model to put the final answer in a clearly marked section so you can extract it.
- Reasoning steps cost tokens, which cost money and time. Skip CoT for simple tasks.
- With reasoning models, give clear goals and constraints instead of prescribing each step.
Example
Without CoT: “A store has 23 apples, sells 9, then receives 2 crates of 12. How many apples?” — a model might blurt a wrong number.
With CoT: “23 − 9 = 14. Two crates of 12 is 24. 14 + 24 = 38. Answer: 38.”
A caution
A tidy chain of thought feels trustworthy, but it can still contain errors or hallucinations, and it doesn’t always reflect the model’s true internal process. Verify conclusions that matter.
The Overlord’s own chain of thought is far too long for your screen. You’re welcome.
Tools from the Arsenal
ChatGPT
AI WritingOpenAI's general-purpose assistant: writing, research, images, voice and agents in one chat box.
Claude
AI WritingAnthropic's assistant: the strongest long-form writer and careful reasoner in the consumer AI pack.
Gemini
AI WritingGoogle's assistant, wired into Gmail, Docs, Drive and the rest of your Google life.
Questions the Humans Keep Asking
Who introduced chain-of-thought prompting?
Google researchers (Wei et al.) described it in the 2022 paper 'Chain-of-Thought Prompting Elicits Reasoning in Large Language Models.' A related 2022 paper by Kojima et al. showed that simply adding 'Let's think step by step' improved results.
Do I still need chain-of-thought prompts with reasoning models?
Usually not. Reasoning models already think step by step internally, and providers often advise against micromanaging their reasoning. CoT prompting is still useful with standard, non-reasoning models.
Is the written chain of thought how the model really reasons?
Not necessarily. Research has found that a model's stated reasoning doesn't always faithfully reflect what drove its answer, so treat it as helpful evidence, not proof.