Chain of Thought
also called CoT, Reasoning out loud
Let the model work through steps before committing to an answer.
Asking the model to reason step by step before giving its conclusion, which measurably improves arithmetic, logic and multi-constraint tasks.
Reasoning models now do this internally, which has changed the advice: for those, prompting "think step by step" manually is usually redundant, and you should ask for the conclusion rather than the transcript.
The advice changed when reasoning models arrived, and it is worth knowing why. Those models generate their reasoning internally as part of training, so manually prompting "think step by step" is largely redundant; what you now choose is whether you want the transcript or the conclusion, and how much thinking budget to pay for.
Reasoning out loud beats reasoning silently
A multi-constraint scheduling problem comes out wrong.
Asking for the working before the answer improves accuracy on constraint-heavy tasks, and it gives you something to debug: the failure is visible in which constraint it dropped. Silent wrong answers tell you nothing.
- paperChain-of-thought prompting elicits reasoning in large language models ↗
Few-shot demonstrations containing intermediate reasoning steps improved results on arithmetic, commonsense and symbolic reasoning tasks.
- docsExtended thinking ↗
Configurable thinking budgets and interleaved reasoning around tool results — the API-level view of chain of thought as a resource you pay for.
- paperTree of thoughts: deliberate problem solving with large language models ↗
Explores multiple reasoning paths with self-evaluation, lookahead and backtracking — search over thoughts instead of one linear chain.
Prompt Engineering
The craft of writing instructions a model reliably follows.
Structured Output · JSON mode
Making the model return data your code can parse.
Evaluation · Evals
How you know the system is actually working.
Instruction Following
The model’s ability to actually obey what you asked.