| Prompt |
Everything given to the model: instructions, context, examples and input |
| Context |
All tokens the model sees when producing one response, including instructions, tool definitions, documents, history and tool results |
| Context window |
Maximum tokens the model can reference at once; whether output and reasoning count in the same window varies by model |
| Token |
The unit in which a model processes text; tokens do not map one-to-one to characters |
| Context rot |
Decline in the ability to accurately recall information in context as token counts increase |
| Context engineering |
Selecting and arranging all the tokens the model sees on each turn |
| System (developer) message |
Position holding roles and rules shared by all conversations; designed to take priority over user messages, but does not fully prevent injection |
| Mid-conversation system message |
A system-authority instruction appended during a conversation, used to change rules while preserving the cache |
| Role prompting |
Adding a sentence defining perspective and tone to the system prompt; supplies neither knowledge nor authority |
| Success criteria |
Observable conditions that let you judge from the result whether it went well |
| Task type |
Nature of work, such as extraction, transformation, generation or judgment, determining permitted invention and required evidence |
| Right altitude |
Level between hardcoded rules and vague principles that provides a judgment framework through principles, reasons and representative cases |
| Prompt injection |
An attack using sentences hidden in input to alter model instructions; direct attacks come from users, indirect ones from third-party content |
| Least privilege |
Granting models and tools only the minimum necessary access to reduce damage if defenses fail |
| Few-shot examples |
Showing a few pairs of inputs and desired outputs; without examples, the approach is zero-shot |
| Example bias |
Skewed model judgment caused by imbalances in example label frequency, order or wording |
| Boundary case |
Input between two categories, or on which people disagree |
| XML tags |
Markers separating instructions, context, examples and input; names are freely chosen but used consistently |
| Data dictionary |
Table of each field’s business meaning, type/unit, allowed values and missing-value representations |
| Structured outputs |
Feature enforcing output compliance with a specified JSON schema |
| JSON mode |
Output mode guaranteeing valid JSON but not schema compliance |
| enum |
List of permitted values, used to constrain classification labels |
| Prefill |
Technique for steering output by supplying the beginning of an assistant message; Claude 4.6 and later do not support last-turn prefill |
| Stop reason |
Why a response ended; Anthropic uses stop_reason, OpenAI Responses API uses status and incomplete_details.reason |
| Citations |
Anthropic feature returning original-source locations with cited portions of a response |
| Search grounding |
Feature supporting model answers with web-search results and sources |
| Retrieval-augmented generation (RAG) |
Retrieving relevant material at question time and putting it in context |
| BM25 |
Word-matching retrieval method strong on exact strings such as numbers and codes |
| Contextual Retrieval |
Indexing chunks with short explanations based on their full documents |
| Reranking |
Reordering retrieval candidates by question relevance and retaining top results |
| Compaction |
Replacing old portions of a long conversation with a summary to continue |
| Context editing |
Clearing old tool results or thinking blocks according to rules |
| Just-in-time loading |
Keeping lightweight identifiers and retrieving content with tools when needed |
| Subagent |
Supporting agent doing focused work in a clean context and returning only a summary |
| Prompt caching |
Storing and reusing repeated request prefixes; exact prefix matches are required |
| Reasoning model |
Model generating internal reasoning tokens before answering |
| Reasoning tokens |
Tokens generated before a reasoning model’s answer, hidden or summarized and billed as output |
| Adaptive thinking |
Model decision on whether and how much to think based on request complexity and effort |
| Reasoning effort |
Setting controlling effort spent on thinking and responses |
| Output limit |
Maximum tokens a request can generate, including reasoning tokens |
| Prompt chaining |
Splitting work into calls, passing each stage’s output as the next stage’s input |
| Routing |
Classifying inputs and sending them to type-specific paths |
| Workflow pattern |
Defined arrangement of calls, such as chaining, routing, parallelization, orchestrator-workers or evaluator-optimizer |
| Self-review |
Pattern with stages reviewing a draft against criteria and revising it |
| Rubric |
Scoring sheet listing per-item pass conditions |
| Critical item |
Evaluation item graded separately from averages because one failure makes the result unusable |
| Test set |
Fixed collection of inputs and expected results for comparing changes |
| Confirmation set |
Holdout used only for final decisions, kept closed during improvement |
| Regression |
Something that worked before a change stops working afterward |
| BLEU / ROUGE |
Metrics scoring word overlap with reference answers; they do not measure meaning or factual accuracy |
| LLM evaluator |
Evaluation method using a model as grader |
| Position bias / verbosity bias |
Evaluator preference for earlier-presented or longer answers |
| pass@k / pass^k |
Probability of at least one success in k attempts, and probability that all k attempts succeed |
| Capability / regression evaluation |
Evaluation of what a system does well, and whether it still does what it could before |
| Transcript |
Complete agent-trial record, including outputs, tool calls and reasoning |