Cheat sheet: Prompting and structured output

Prompting types:

Type	When
Zero-shot	Simple, well-known task
Few-shot	Need a specific output format or style
Chain-of-thought	Multi-step reasoning on a model that does not already reason
Self-consistency	High-stakes reasoning with a short answer; cost is about N times

Structured output options:

Option	Guarantees
Prompt for JSON, then parse	Nothing
JSON mode	Valid JSON, not your schema
Schema-constrained outputs	Conforms to your schema; not truth, completeness or refusals
Tool calling	Schema-checked arguments in strict mode; for actions, not a “stronger” option

Quick answers

Reasoning models: spend inference-time compute; reasoning tokens are billed as output; use for hard multi-step tasks; cost latency and tokens; visible reasoning is not an audit trail.
Structured output: constrain, then validate (form and meaning), bounded retry with the error, fallback. Valid does not mean correct.
Prompts: delimit untrusted text; judge changes on a labelled set and pass rate.
Check the docs: effort and thinking settings and supported schema features differ by model and change often.