Glossary

Each term gets one sentence. The line under it is the common mix-up.

• Model. The AI engine that turns your instructions and context into an answer. It is only one ingredient in the result.
• Prompt. The request you are making right now. A good one states the outcome you want, not just the topic.
• Context. Everything the AI can see while it works on your request. The right material helps. More material does not.
• Token. A small piece of text. Models read and write in tokens, and billing is counted in them. You will not count them. Just know that bigger jobs cost more.
• Tool. A capability the AI can call to look something up or take an action, such as search or a database query. Tools give it power, so they need permission limits and checking.
• Connector. A link between the AI and a real system such as Jira, Slack, or Drive. Being connected does not mean it should touch everything inside.
• MCP. An open standard for plugging tools and connectors into AI. It is plumbing. It does not make the AI smarter or safer.
• Skill (or recipe). A saved procedure the AI can run again for a job you repeat. Products call it a skill, command, or template. It is the same idea.
• Agent. AI that picks its own next steps toward a goal instead of giving one answer. More freedom is not automatically better than a plain chat.
• Workflow. A fixed path of steps, checks, and handoffs that runs the same way every time. Repeatable value comes from this, not from one lucky answer.
• Eval. A repeatable test of whether an AI setup does a job well enough. Five of your own real examples beat a public leaderboard.
• Benchmark. A public test used to compare models. Treat a score as marketing until you know how it was produced.
• Spec (or durable brief). A written record of what you want, why, and what was decided. It lets the work outlive the chat and the person who started it.
• Hallucination. When AI states something false with full confidence. It is not lying. It is guessing. That is why you check.