Plain-language definitions for the terms that matter in practice. Every entry links to the lesson
that teaches it.
- Action item
A single commitment from a meeting: one action, one owner, one date — or an explicit 'unassigned' / 'no date' where the …
- Caveat
A condition or limit attached to a statement — the 'if', 'unless' or 'to be confirmed' that decides when it is true. Sum …
- Context window
The context window is everything a model can consider at one time: your current message, the conversation so far, attach …
- Data classification
Sorting content before it travels: public, internal, confidential, personal, secret. Each level has its own rule for a t …
- Dry run
Running a script or process in a mode that only reports what it would do, without changing anything. The habit that turn …
- Few-shot prompting
Few-shot prompting means showing the model two or three input → output examples of the pattern you want, then giving it …
- Generative AI
Generative AI is software that produces new content — text, images, audio or code — by continuing patterns learned in tr …
- Grounding
Grounding means constraining a model's answer to material you supply — a pasted policy, paper or document — instead of i …
- Hallucination
A hallucination is a fluent, confident AI answer — or a detail inside one — that is not true. Because models generate wh …
- Large language model (LLM)
A large language model is an AI system trained on enormous amounts of text to predict the next token in a sequence. When …
- Placeholder
A marked gap in a draft — like [refund reference] or [amount] — that only you can fill with a real fact. Asking for plac …
- Primary source
The original artifact a claim comes from — the study itself, the law's text, the official documentation, the dataset. Se …
- Prompt
A prompt is everything you give a model to work with: the instruction, the context, the examples and any material you su …
- Prompt injection
Instructions hidden inside content you ask an AI to process — a document, web page or email — aimed at the assistant its …
- Reasoning model
A reasoning model spends extra computation before answering — working through steps, checking itself, exploring alternat …
- Redaction
Removing or replacing sensitive values — names, keys, amounts — with safe stand-ins before text travels anywhere it coul …
- Rubric
A short list of checkable questions you run against a draft before shipping it — is the format right, can every claim be …
- Structured output
Output shaped to a fixed format — a table, a list of fields or JSON — so it can be reused by other tools instead of re-t …
- Temperature
A sampling setting that controls how randomly a model picks its next token: lower values make output more predictable, h …
- Token
A token is a chunk of text — roughly a word piece — that a model processes as one unit. English averages about three-qua …
- Training cutoff
The date after which a model's training data stops. Anything newer — prices, laws, product versions, news — is beyond it …
- Triage
Sorting incoming items — messages, tickets, requests — into defined categories and a priority order before working on an …
- Unsupported claim
A statement specific enough to be checked and consequential enough to matter — attached to no evidence. Precise numbers, …