Plain-language definitions of the AI terms used across our free tools — no prior AI knowledge assumed.
AI models read and write text in small chunks called tokens — roughly ¾ of a word. 1,000 tokens ≈ 750 words. You're billed by the token, which is why this number matters.
Shows up in: AI Cost Estimator, Token Counter, Model Picker, Prompt Auditor, Prompt Caching Savings Analyzer, Batch API Suitability Checker, Context Window Budget Planner, Cost-Aware Routing Designer, Rate Limit & Capacity Planner, RAG Chunking & Retrieval Planner
Prompt caching lets an AI model reuse the instructions and context it saw on a recent request instead of processing them again — like not re-reading a manual you just read. Repeated (cached) tokens are billed at a lower rate than fresh input tokens.
Shows up in: AI Cost Estimator, Prompt Auditor, Prompt Caching Savings Analyzer
A cached prompt doesn't stay cached forever — its TTL (time-to-live) is how long it survives with no calls before it expires and has to be written again. Anthropic's default is 5 minutes, with a 1-hour option at a higher write price; each reuse resets the clock, so a prompt called more often than its TTL effectively never falls out of cache.
Shows up in: Prompt Caching Savings Analyzer
The Batch API processes requests in bulk instead of instantly, in exchange for a lower price. It fits things like overnight report generation or bulk classification — not a live chat where someone's waiting on screen.
Shows up in: AI Cost Estimator, Batch API Suitability Checker
A system prompt is the fixed instructions you send an AI model on every request — its job description. It usually stays the same across many requests, which is what makes it eligible for prompt caching.
Shows up in: Prompt Auditor, Context Window Budget Planner, Prompt Optimizer
The context window is the total amount of text (measured in tokens) a model can "see" at once — its instructions, the conversation so far, and anything you've given it to read. Some models handle very long input (a whole codebase, a long contract) better than others.
Shows up in: Model Picker, Context Window Budget Planner
An eval is a set of example inputs with a description of what a correct response looks like, run against your AI feature to check whether it works before and after a change — like unit tests, but for AI behavior instead of code.
Shows up in: Eval Set Generator
A model cascade tries a cheap, fast model first and only calls a stronger (more expensive) model when the cheap one's answer looks wrong — measured by a confidence score, a failed validation check, hedging language, or a hard input-complexity rule. It only saves money if the cheap model clears the bar often enough; escalate too often and the cascade costs more than just using the strong model for everything.
Shows up in: Cost-Aware Routing Designer
Anthropic caps how much you can call the API per minute in two ways: RPM (requests per minute) and TPM (tokens per minute). Your limit depends on your account's tier and is per-account, not published like model pricing — check the Anthropic Console to see yours. Exceeding either one gets a request rejected (HTTP 429) until the next minute rolls over.
Shows up in: Rate Limit & Capacity Planner
MCP is how Claude connects to external tools and data — like giving it hands to use a calendar, a database, or a file system. Different Claude apps expect this connection info in a different config file, in a different place on disk.
Shows up in: MCP Config Wizard
A subagent is a separate, specialized Claude Code configuration (a name, a trigger description, a scoped list of tools it's allowed to use, and its own instructions) that the main session can delegate a task to. Claude Code decides whether to use one based on how well its description matches the current situation — the same triggering concern a skill file has.
Shows up in: Subagent & Slash Command Generator
Tool use (also called function calling) lets a model call code you wrote — looking up a real order status, running a calculation, searching a database — instead of only generating text. The model decides WHETHER and HOW to call your tool almost entirely from the description and parameter descriptions you give it, which is why a vague description is the most common cause of a tool being called at the wrong time or with the wrong arguments.
Shows up in: Tool / Function Schema Generator
JSON Schema is a standard way to describe the exact shape a JSON object must have — which fields exist, their types, which are required, and constraints like allowed values or numeric ranges. Giving a model a JSON Schema (and validating its output against one) is the most reliable way to get structured output you can trust without eyeballing every response.
Shows up in: Structured Output Schema Builder, Tool / Function Schema Generator
Few-shot prompting means showing a model 1-3 worked examples of the task instead of (or alongside) describing it in words. It reliably helps for subtle formatting or edge-case behavior clear instructions struggle to convey, but every example is paid for on every single call — and a set of examples that all happen to share an irrelevant trait can teach the model that trait instead of the task.
Shows up in: Few-Shot Example Set Builder
RAG lets a model answer from a large corpus it wasn't trained on: your documents are split into chunks and embedded in advance, then at query time the most relevant chunks are retrieved and inserted into the prompt as context. Getting chunk size, overlap, and how many chunks to retrieve (top-k) right is most of the tuning work — get them wrong and the model answers confidently from the wrong or missing context.
Shows up in: RAG Chunking & Retrieval Planner