Glossairy — AI Language, Clarified

Basics

Agent

agentic

A model given tools, memory, and a goal — then left to reason through the steps. The leap from answering questions to completing work.

Browser use

A model that navigates the web like a human — clicking, scrolling, filling forms. Turns any website into an API.

Computer use

A model that controls a computer directly — mouse, keyboard, screen. The most general-purpose form of agentic capability.

Context window

The amount of text a model can hold in working memory at once. Everything outside it doesn't exist.

In context learning

AGENTS.md
CLAUDE.md
ICL
memory

Teaching a model by example within the prompt itself — no retraining required. The basis for few-shot prompting and persistent instructions.

Model

foundation model
language model
LLM

Software trained on data to recognize patterns and generate outputs. The core building block of modern AI.

Open

free
open source
open weight

Models whose weights are publicly available. "Open source" is often used loosely — open weight is more precise.

Prompt

The input you give a model. Every interaction starts here. Writing them well is the highest-leverage skill in AI.

Reasoning

chain-of-thought
thinking
understanding

Models produce outputs that look like reasoning. Whether they actually reason is an open debate.

Structured

JSON
schema
structured output

Output formatted to a predictable shape so software can reliably parse and use it.

Tools

function calling
MCP
tool use

The ability for a model to call external functions — search, code execution, APIs — and use the results.

Dichotomy

Branding

ChatGPT

The product. A consumer interface built by OpenAI.

OpenAI

The company. Conflating them is like saying "I use Google" when you mean Gmail.

Cognition

Language

Models are trained on language and produce fluent text.

Logic

Fluent text isn't the same as correct reasoning. Models can produce plausible conclusions that are mathematically wrong.

Data

Structured data

Databases and spreadsheets. Predictable shape, easy to query.

Unstructured data

Emails, documents, conversations. Models are uniquely good at this.

Execution

Agentics

Prompts and models — adaptive, probabilistic, capable of handling the unexpected.

Automation

Loops and scripts — predictable, brittle, deterministic.

Interface

Graphical

Most software is built around visual interfaces designed for human interaction.

Textual

AI-native software is often built around text — which changes everything about how you design it.

Retrieval

LLM

large language model

Generates from memory alone.

RAG

grounding
retrieval-augmented generation

Retrieves relevant documents first, then generates — reducing hallucination at the cost of complexity.

Software

behavior

Deterministic

Code runs the same way every time.

Probabilistic

The same prompt can produce different outputs. This is a feature and a bug.

Training

Supervised

Trains on labeled examples. The model learns to predict correct answers.

Unsupervised

Finds patterns without labels. Most modern LLMs use both at different stages.

Pathology

Cargo cult

best practices
prompt engineering

Copying AI workflows or prompts without understanding why they work — or whether they do.

Disingenuous selection

demo
it just works

Only showing what works. It can take 1,000 attempts to get one result worth showing. Presenting that result without disclosing the fail rate is technically true and 99.9% deceptive.

AI content
generated output

When someone generates text they expect you to read but haven't read themselves.

Fidelity of desire

Do you really want an AI running your life? Adults make 35,000 decisions per day. Offloading them feels efficient — until you realize presence was the point.

Intent preservation

Ensuring a model does what you actually meant, not just what you literally said. The gap between instruction and execution is where most failures live.

Mingentic

The minimum viable agentic unit. A focused model call with a clear input, constrained scope, and defined output contract.

Ontological self

The stable behavioral persona a model develops through training. Not consciousness — but a consistent point of view that shapes every output.

Rapid monologue

A model generating a long uninterrupted stream before acting. Trades responsiveness for depth — and depth for hallucination risk.

Superwork

superworker

Work accelerated and expanded by agentics. Not automation — judgment and tools working together toward outcomes that weren't possible before.

Unicorn

A person who can design, build, and ship. Rare before AI. Less rare now.

Vibe coding

Building software by describing what you want and letting a model write it. The craft shifts from syntax to taste.

Virtual employee

virtual workforce

An agent with a defined role, responsibilities, and tools — not just a one-off task runner. The difference between a contractor and a colleague.