Glossary
The shared vocabulary of an AI-fluent team.
From RAG to MCP to context windows — plain-English definitions for the terms we use every day, so nobody has to nod and pretend to know.
- .env
- A plain text file (named literally
.env) that holds the API keys and other secret values a script needs to run. Lives outside the code so credentials never end up in GitHub or pasted into Claude. - .md
- Short for Markdown — a lightweight text format where simple symbols like
#,*, and-mark up headings, emphasis, and lists. Most AI prompts, agent definitions, and docs at PDG live in.mdfiles. - Agent (Claude)
- A reusable Claude configuration with its own role, instructions, and tools — built to handle a specific job (a recruiter agent, a research agent, a personal assistant). You “open” an agent the same way you’d open a folder.
- API
- Application Programming Interface. The plug a piece of software exposes so other software can talk to it. When we say “calling the Claude API,” we mean sending Claude a prompt programmatically instead of typing into the chat window.
- Artifact (Claude)
- A self-contained file Claude generates inside a conversation — a doc, a chart, a snippet — rendered in a panel next to the chat so you can edit, download, or share it.
- CLAUDE.md
- A markdown file in a project folder that tells Claude how to behave in that project: the rules, the file map, the working style. When you open the folder in Claude Code, this file loads automatically as context.
- Command (Claude)
- A slash-prefixed shortcut in Claude (e.g.,
/ultrareview) that runs a predefined workflow. Faster than re-typing the same instructions every time. - Connector (Claude)
- A pre-built integration that wires Claude into an outside service — Google Drive, Gmail, Slack, GitHub — usually through a one-click sign-in. The productized, no-setup version of MCP for everyday users.
- Context window
- The total amount of text Claude can hold in working memory at once — your prompt, attached files, the running conversation, and Claude’s reply. Once it fills up, older content has to be summarized or dropped.
- Hallucination
- When an AI confidently produces something that sounds right but isn’t — a made-up citation, a fake quote, a wrong number. Why human review is non-negotiable on anything that ships.
- LLM
- Large Language Model. The kind of AI Claude, ChatGPT, and Gemini all are: trained on huge volumes of text to predict and generate language. When you “talk to Claude,” you’re talking to an LLM.
- Live Artifact (Claude)
- An artifact that runs interactively — a working web app, a calculator, a chart you can poke at — rendered live inside the Claude conversation rather than handed back as a static file.
- Local
- Running on your own machine, not in the cloud. “Run it locally” means the code is executing on your laptop; “local files” means files saved to your hard drive, not Google Drive.
- MCP
- Model Context Protocol. The open standard that lets Claude plug into outside tools and data sources — Asana, Drive, Slack, a database — so it can read and act on real information instead of just chatting. Sanctioned at PDG only inside Claude Enterprise.
- OAuth
- The standard “Sign in with Google / Sign in with Microsoft” flow. Lets an app act on your behalf without ever seeing your password — it gets a scoped token instead. How Claude Enterprise authenticates through PDG SSO.
- Plugin (Claude)
- A bundle of skills, commands, and agents you install into Claude to extend what it can do. Think browser extensions, but for Claude.
- Prompt
- What you type to Claude. The instruction, the question, the brief. The better the prompt, the better the output — same as briefing a person.
- RAG
- Retrieval-Augmented Generation. A pattern where Claude looks up relevant documents from a private library before answering, so its response is grounded in your real material instead of just its training data.
- Skill (Claude)
- A reusable bundle of instructions and rules Claude can load on demand to handle a specific task — formatting a doc, running a code review, building a frontend. Like a plug-in skillset Claude can pick up and put down.
- Token
- The unit AI models count text in — roughly three-quarters of a word in English. When you hear “context window of 200K tokens,” that’s about 150,000 words. Pricing and limits are quoted in tokens.
- WAT framework
- Workflows, Agents, Tools. PDG’s internal pattern for building AI systems: probabilistic AI (agents) handles reasoning, deterministic Python (tools) handles execution, and markdown workflows define the steps.