| Term | Easy meaning |
|---|
| Prompt engineering | Designing prompts so the model works reliably, not just “asking nicely”. |
| Structured output | A response in a strict shape, often JSON, that software can safely parse. |
| Reasoning effort / thinking level | How much internal thinking compute the model uses before answering. |
| Grounding | Using search, retrieved docs, or tools so answers are based on evidence. |
| RAG | Retrieval-Augmented Generation. Search relevant documents and give them to the model. |
| Prompt injection | When untrusted text tries to act like instructions and hijack the model. |
| Planner–executor | One step or model plans; another step or model executes the work. |
| Evaluation flywheel | Analyze failures, measure them, improve prompts, test again. |
| Lost in the middle | Important facts in the middle of a long prompt are easier to miss. |
| Thought signatures | Gemini mechanism for preserving reasoning context in some flows. |
| Done criteria | Clear definition of what finished, correct work looks like. |