Prompt engineering specifies the task, context, output contract, and checks used when a model produces work. Reliability comes from testable boundaries, not a longer instruction by itself.
Choose a starting point
Begin with a narrow task and representative inputs. Add examples or retrieved context only where they reduce ambiguity; then define structured output, failure handling, evaluation cases, and a review path for uncertain answers.
- Prompt Engineering
- Prompt engineering: define a task that can be checked
- Zero-shot baseline: measure the task before adding demonstrations
- Prompt intent: turn an open request into an acceptance check
- Prompting, retrieval, or training: choose the failing layer
- Prompt failure diagnosis: find the broken contract first
Common Mistakes
Prompt text cannot grant a model access to missing facts or make an untrusted document authoritative. The linked topics distinguish useful context from injection, leakage, and unsupported certainty.