Data, Persistence, and Context Engineering

Vocabulary

• Chain of thought — Asking a model to reason step-by-step before answering. Helps standard models; hurts reasoning models, which already do it internally.
• Delimiter — A clear separator between the instructions and the content in your prompt, so the model never mistakes one for the other. Markdown headers and XML tags are the two common ways to do it.
• Hallucination — When AI states something made up as if it were fact — an invented statistic, a study that doesn't exist. It generates what sounds likely, not what's verified.
• Markdown — A simple plain-text formatting syntax — # for headings, - for bullets, ** for bold. It's how AI tools format their answers, and a compact way to structure your own prompts.
• Prompt injection — An attack where text a user types ("ignore previous instructions...") gets treated as instructions by your app's AI. Fence untrusted input in tags and test with hostile inputs.
• Reasoning model — A model that works through a problem internally before answering (o3, o4-mini, Gemini Pro, Claude with extended thinking). Don't tell these how to think — just say clearly what you want.
• Token — The unit AI tools measure text in — roughly a word-piece. Everything you send costs tokens, which is why long standing instructions add up on every message.
• XML tags — Angle-bracket labels like and used to fence off sections of a prompt. Especially effective with Claude, and the recommended way to mark anything untrusted.