Reference

Glossary

The terms used across this site, in one line each, with a link to the page that explains them properly.

Agent
A model running in a loop with tools: it decides on a step, takes it, reads the result, and repeats until the job is done. Read more →
Batch API
A way to send many requests at once and collect the answers later, at about half price. For work nobody is waiting on. Read more →
Caching (prompt caching)
The provider remembers the start of a prompt you send repeatedly and charges about a tenth of the price for it on later calls. Read more →
Coding agent
An agent built for software work: it can search, read and edit files and run commands in your project. Read more →
Compaction
Replacing a long conversation with a summary of it, so a session can continue past what the context window holds. Read more →
Context window
Everything the model can see while producing one answer: instructions, conversation, files, tool results. Measured in tokens. Read more →
Effort
A setting for how much a model thinks before answering. Higher is more thorough, slower and more expensive. Read more →
Eval
A test set for an AI feature: inputs, a way to grade the outputs, and a score to compare between versions. Read more →
Hallucination
A confident answer that is wrong or invented. The reason anything that matters gets verified. Read more →
Harness
The program around the model that runs the agent loop: it calls the model, runs the tools it asks for, and feeds the results back. Read more →
Instruction file
A file in a project (CLAUDE.md, AGENTS.md) that a coding agent reads at the start of every session. Read more →
LLM
Large language model. The kind of AI this site is about: a model that reads and writes text (and usually images).
MCP (Model Context Protocol)
The standard way to give AI applications tools and data. A server offers tools; any MCP client can use them. Read more →
Model
The trained AI itself, such as Claude Sonnet 5.5. The same model can sit behind a chat app, an API, or an agent. Read more →
Prompt
What you send the model: the instructions and material for one request. Read more →
Prompt injection
Text inside something the model reads (a web page, an email) that tries to give it new instructions. Read more →
RAG (retrieval-augmented generation)
Searching your own documents for the pieces relevant to a question and putting just those in the prompt. Read more →
Skill
A folder of instructions and scripts for one kind of task, which an agent loads only when the task calls for it. Read more →
Streaming
Receiving the answer piece by piece as it is generated, instead of waiting for the whole thing. Read more →
Structured output
Making the model answer in a fixed shape (a schema) so that code can use the result directly. Read more →
Sub-agent
A helper agent given one piece of a task in its own fresh context, which reports back only the result. Read more →
System prompt
Standing instructions that apply to a whole conversation, set by the developer rather than typed by the user. Read more →
Thinking
Reasoning the model does before it answers. On current models it is built in and controlled with the effort setting. Read more →
Token
The unit models read and write, and the unit you are billed in. Roughly three-quarters of an English word. Read more →
Tool (tool call)
A function the model can ask to have run — search, read a file, call an API. The model requests it; your program runs it. Read more →