Few-shot prompts add worked examples and reasoning prompts ask for intermediate stages. Both change answers, token cost, and failure modes, with effects that vary by model and must be measured.
Term | Meaning |
|---|---|
Zero-shot | A prompt with instructions and the new input, and no worked examples. |
Few-shot | A prompt with a small set of worked input-output pairs before the new input. |
Reasoning trace | Intermediate stages shown before the final answer, such as partial sums. |
Self-check | A follow-up instruction that asks a model to re-examine its own answer. |
Structured output | An answer in a fixed machine-readable shape, such as fixed keys and types. |
Lesson | What it covers |
|---|---|
What examples change, with a worked cost comparison. | |
Coverage, order stability, format, and context budgeting. | |
Where stages help, what they cost, and their limits. | |
Re-reading requests, fixed output shapes, and independent verification. |
Whether the gain holds across inputs and models. One trial is not evidence, since effects vary by model and must be measured.