---
name: ai-coding-terminology-precision
description: "Precise AI-coding terminology for docs and technical discussion. Use when discussing AI-assisted coding concepts."
---

# AI Coding Terminology Precision

- Use AI-coding terms precisely in documentation and discussion; the dictionary exists because vague jargon obscures real mechanics.
- Token: the atomic unit the model reads and writes, roughly word-sized; context size and cost are measured in tokens.
- Model: parameters produced by training; stateless; it only performs next-token prediction and carries nothing between requests.
- Context window: everything the model sees for one request, as a single finite token sequence - not a database, not memory.
- Session: one bounded interaction run; stateful across turns but not across sessions. Cross-session persistence requires explicit memory or handoff artifacts.
- System prompt: standing instructions prepended to every request that shape agent behavior; distinguish it from user messages when diagnosing behavior.
- Tool: a function exposing perception of or action on the environment to the agent; describe agent capability in terms of its tools.
- Hallucination: confidently wrong output; distinguish factuality gaps (missing knowledge) from faithfulness drift (contradicting loaded context).
- Do not call "AI" a fixed technology in docs; it is a moving label - name the actual mechanism (model, agent, tool call) instead.
- When explaining cost, note it scales with requests, not user turns: one user message can trigger many tool-calling requests.