Short, plain definitions of the terms you will meet when choosing an AI tool. Written by TaskRated.
- Agent
- An AI system that can take steps on its own, such as searching, opening files or using other software, to finish a task rather than only replying to a message.
- Agent skill
- A packaged set of instructions (and sometimes scripts) that teaches an AI agent how to do one kind of task. Often shared as a folder with a SKILL.md file.
- Context window
- How much text a model can consider at once, including your prompt, files and its own reply. Long documents may be cut off if they exceed it.
- Fine-tuning
- Further training a model on examples so it behaves a certain way. Most small businesses do not need it; good prompts and examples usually suffice.
- Free tier
- A no-cost plan, usually with limits on messages, features or exports. Results on free tiers can differ from paid plans.
- Grounding / citations
- Linking an answer to sources. A citation is only useful if the page exists and actually says what is claimed, so open the links.
- Hallucination
- When a model states something false as if it were true, such as an invented statistic, quote or skill. Our tests plant traps to catch this.
- Large language model (LLM)
- The kind of model behind chat assistants. It predicts likely text, which is why it can sound right while being wrong.
- Model
- The underlying AI system a product uses. One product may offer several models with different speed, cost and quality.
- Prompt
- The instructions and data you give an AI tool. Clear role, task, data and output format usually beat clever wording.
- Prompt injection
- Hidden instructions inside a document or web page that try to make an AI tool do something you did not ask. A risk when agents read untrusted content.
- RAG (retrieval-augmented generation)
- Having the tool look up relevant documents first and answer from them, instead of only from memory.
- System prompt
- Background instructions set by the product or by you that shape every reply, such as tone or rules.
- Token
- The unit models count text in, roughly three-quarters of an English word. Limits and prices are often quoted in tokens.
- Usage limit
- A cap on how much you can use a tool in a period. Hitting it mid-task counts against friction in our scores.