Plain English Glossary.
A dictionary of every term you'll hear in the workshop. Plain-speak definitions, no jargon, no homework. The goal isn't to memorize it. The goal is to know where to look when something sounds like a foreign language so you can keep moving through the build.
- A working vocabulary for the workshop's core ideas, the tools we install, and the patterns we build on top.
- Plain definitions for the words that sound technical (MCP, sub-agent, heartbeat, cron) so you can read any other card without stalling.
- A reference you'll return to every time a new term lands in chat or a guide and you want to check what it actually means in this workshop.
How this fits in the OS.
Nothing. Open it before you start, keep it open while you build, come back any time a word in another card sounds like a foreign language.
The actual vocabulary the workshop uses, pulled from the guides and the rest of the library. Same words, plain definitions.
Reading any other card without stalling on a term. Once the words click, the patterns click. Once the patterns click, the builds click.
How to read this card.
Four groups, browse the one that fits your question. A. Core vocab is the words you'll hear in the first ten minutes (agent, prompt, context, model). B. Tools and infrastructure is the moving parts the workshop installs (CLI, skill, MCP, slash command, sub-agent). C. Knowledge architecture is the second-brain layer (cairns, L1/L2/L3, vault, vector search). D. Workflow modes and commands is the safety and automation vocabulary (safe mode, auto mode, /personalize).
Every entry is short on purpose. When a word has a longer card, the entry points to it. Where Claude Desktop uses a different name for the same thing, the entry says so.
A. Core vocab.
The words you hear before any tool gets installed.
The basics.
An AI assistant that can actually do things on your computer, not just chat about them. It reads files, runs commands, calls tools, and edits documents on your behalf. Claude Code is an agent. The desktop Claude chat is closer to a chatbot.
A way of working where the system is built for agents to use, not just humans. Plain text files instead of locked databases. Clear file paths. Predictable structure. The workshop title is about this shift: stop treating AI like a helper inside human tools, start treating your computer like an operating system that agents and humans share.
Anthropic's family of AI models. The brand name. When someone says "Claude," they usually mean a specific surface (the chat website, the desktop app, or Claude Code in the terminal). The model underneath is one of Opus, Sonnet, or Haiku.
The agent that runs in your terminal. The most powerful way to use Claude in this workshop, because it can read and write files, run commands, connect to your tools, and act on its own. See What Is Claude Code.
The software wrapped around an AI model that turns it from a thing that writes text into an agent that can act. It runs the loop (think, act, read the result, repeat), gives the model real tools (your files, the terminal, the web, your connected apps), manages its context, and stays inside the permissions you set. Claude Code is a harness: the model is the brain, the harness is everything that lets it do things on your computer.
Everything the agent currently knows in a session. What you typed, what it read, what tools it called, what it found. Context lives in a context window with a hard size limit. When that fills up, quality drops. Most of the patterns in this workshop exist to manage context.
The working memory of a single agent session, measured in tokens. Once it's full, the agent starts losing the earliest things you told it. You feel that as "Claude got dumber three hours in." Sub-agents and the Cairns memory tiers exist to keep this window lean.
The unit the model and your bill are counted in. Roughly three quarters of a word in English. You don't need to count them. You just need to know "more tokens" means slower and more expensive, which is why we keep agents small and focused.
What you type to the agent. Could be one sentence, could be a whole document. The starter prompts on these cards are pre-written prompts you can paste straight into Claude Code.
The standing instructions an agent reads before anything you type. Your CLAUDE.md works like a system prompt for your main agent. Sub-agents get their own.
The actual AI underneath. Anthropic ships three tiers: Opus (most capable, most expensive, used for hard reasoning), Sonnet (the daily driver), Haiku (cheap and fast, used for bulk reading and worker sub-agents).
Claude Code's safety net for a full context window. It automatically summarizes earlier parts of your conversation to free up room. It works, but it's lossy. Spawning sub-agents is the proactive version of the same fix, see Sub-Agents and Agent Teams.
When the model makes something up that sounds right. Made-up citations, invented function names, plausible but wrong facts. The fix isn't trusting harder, it's giving the agent access to real sources (your vault, the web, your files) so it has something to ground its answers in.
A plain text file that lives in your workshop folder. Claude Code reads it automatically every session, so the agent already knows who you are, what you're working on, and how you want to be talked to. See Your Brain File.
The folder on your computer (named agent-native-os) that holds the workshop's guides, templates, and your CLAUDE.md. When the guides say "in your workshop repo," this is the folder they mean.
B. Tools and infrastructure.
The moving parts of the system. Each one has a card that goes deeper.
What gets installed.
A text-based chat window with your computer. You type a command, the computer responds. Claude Code lives here. See Terminal Basics.
The program inside the terminal that actually reads your commands. On Mac it's usually zsh. On Windows you're either in Git Bash or in WSL2's Ubuntu shell. All three accept the same handful of commands the workshop uses.
Command-line interface. A tool you run by typing in the terminal instead of clicking buttons. Claude Code is a CLI. The N8N Toolkit, Slack Toolkit, and Google Workspace CLI are also CLIs. See Borrow a Working CLI and Build Your First CLI.
A small program on your computer that gives Claude a door out to one external service (Gmail, Calendar, Notion, Slack, GitHub). You install it once, the agent can use that service forever. Claude Desktop calls these connectors. Same idea, different name. See MCP Servers.
Claude Desktop's name for an MCP server. If a doc or video mentions a "connector," translate that to "MCP server" in your head.
A reusable capability the agent automatically reaches for when the situation matches. A markdown file at ~/.claude/skills/<name>/SKILL.md with a description telling Claude when to use it. You don't trigger a skill by name. You describe the situation, Claude recognizes the match.
A named shortcut you trigger explicitly by typing /name in Claude Code. A markdown file at ~/.claude/commands/<name>.md. Same idea as a skill, different trigger. You pick a slash command when you want to be the one starting the workflow. See Slash Commands and Skills.
A spawned copy of the main agent, sent off to do one focused task in its own context window. The main agent waits for the result, then keeps going. The whole point is to keep your main context lean by pushing heavy reading down to a sub-agent that throws away its context when it's done. See Sub-Agents and Agent Teams.
A sub-agent that runs in parallel while your main thread keeps going. You don't wait for the result inline, you check on it later. How you do overnight work without sitting at the keyboard.
A coordinated group of sub-agents working on a single combined output, usually with one acting as orchestrator and the others as specialists or workers. The student-facing way of saying what the build-team calls a "fleet."
The build team's name for their production agent-team setup: many sub-agents and background agents running across several machines, with always-on monitoring and scheduled jobs. Workshop students don't run a fleet. The student-side equivalent is an agent team plus a couple of scheduled commands, see Heartbeats and Crons.
A loop that fires every N minutes regardless of the clock. The agent checks something, decides if anything changed, acts only if it has to. Use for "watch for new email" or "alert me if a job fails." See Heartbeats and Crons.
A scheduler that runs a command at fixed clock times: 7:30 AM weekdays, Friday at 5 PM, Monday at 9 AM. Mac and Linux have it built in (crontab -e). On Windows the equivalent is Task Scheduler. Use for "morning brief" or "weekly summary."
A workflow automation platform you self-host. Think of it as the multi-step glue between your tools when a single cron line isn't enough. The workshop uses it for anything that involves triggers, conditions, or multi-step flows across services.
A web-scraping service the workshop uses to pull clean markdown out of any URL or website. The default plumbing for the first phase of building any new CLI.
Secret managers. Tools that hold your API keys and let the agent reach them by reference (op://Vault/Item/credential) instead of pasting the real value into a config file. 1Password is the workshop's recommended path for solo work, Infisical is the team-scale path.
A short string of characters that proves to a service (Gmail, Notion, Brave Search) that you're authorized to use its API. Treated like a password. Lives in 1Password, never in a chat or a committed file.
C. Knowledge architecture.
The second-brain layer and the patterns the workshop uses to keep it agent-readable.
Memory, knowledge, and retrieval.
The folder where your notes live. The workshop uses Obsidian, which is just plain markdown files in a folder Claude can read and write. Not locked in a company's cloud database. Yours, on disk, forever.
A free note-taking app that opens a folder of markdown files. The workshop's default second-brain tool because the files stay plain text the agent can read directly. Plugins worth knowing: Templater, Dataview, Calendar, Omnisearch.
The workshop's name for the three-tier memory pattern (L1, L2, L3) the build team uses for its own knowledge base, and that you'll build a student version of in Stage 02. Named after the stacked-stone trail markers you follow up a mountain when the path isn't obvious. See The Cairns Three Tiers.
The top Cairns tier. A small set of always-loaded waypoints, usually a few markdown files, that tell the agent what matters right now: active projects, key decisions, current voice rules, who's in your life. Always in context. The first stop for any question.
The middle Cairns tier. One dense summary card per real source (a YouTube transcript, a meeting, a paper, a Slack thread). Each card is small and searchable, with a link back to the raw source. The agent reads cards to decide what to open next.
The bottom Cairns tier. The unprocessed source material: full transcripts, PDFs, exports, recordings. Never edited by an agent. The agent only opens an L3 file when an L2 card pointed to it and the question needs exact details.
The technique used to write a good L2 card. You write a short summary, list what's missing, rewrite at the same length but denser, repeat until the card is small but specific. The goal isn't a longer summary, it's a better one. See Chain-of-Density Cards.
A small local agent that knows your memory tiers and decides which one to query first. Keeps your main context lean by handling retrieval as a separate, mediated step. See The Librarian Agent.
The pattern of pulling relevant snippets from a knowledge base and handing them to the agent before it answers. Cairns is one way to do RAG. "Agentic RAG" just means the agent itself decides what to retrieve and when.
A way to search by meaning instead of by exact words. Each chunk of text is turned into a list of numbers (an embedding) and queries find chunks with similar numbers. How an agent finds "the section about pricing" even when the document never uses the word "pricing."
The list of numbers that represents a chunk of text inside a vector search system. You don't read embeddings directly. The agent uses them under the hood to find related content in your vault.
A note-organizing structure (Projects, Areas, Resources, Archive) from Tiago Forte. The default vault layout in the Obsidian guide. Good for humans browsing. Cairns is the structure on top for agent retrieval. You can use both.
D. Workflow modes and commands.
How Claude Code behaves while it works, and the named workshop commands that ship with the repo.
Safety and automation.
Claude tells you before it does anything and waits for your approval. The recommended starting point. The workshop calls this Safe Mode, and the /safe command snaps you back to it; some also call it "Director's Cut."
The agent researches and writes a plan, but doesn't execute. Useful before any big multi-step build. You read the plan, edit it, then tell the agent to run it.
The agent decides what is safe to run without asking, with safeguards against prompt injection. A rung up from Ask permissions; sometimes called Co-Pilot. Picking it the first time in a folder asks you to confirm once. Move here once you trust the workflows you use most.
The agent takes action with no prompts at all. The most autonomous setting, and switched off until you enable it in Claude Code settings. Sometimes called Autopilot or YOLO. Reserved for workflows you've already run dozens of times and trust completely.
The five rungs of how the agent fits into your day: (1) you ask, Claude does; (2) you trigger a command, Claude executes; (3) scheduled, Claude executes; (4) event-triggered, Claude executes; (5) multi-agent, continuous. Most students move from rung 1 to rung 3 in one afternoon.
Per-tool approval rules that decide what an agent can do without asking. Read this file, yes. Edit this folder, ask first. Run this command, never. Live alongside autonomy mode, set separately for sub-agents.
A slash command in the workshop repo. Twenty-one questions about you, your work, and how you learn. Claude turns the answers into a CLAUDE.md tailored to you. Run once before the workshop. See The /personalize Flow.
A slash command that pulls the latest version of the workshop repo and walks you through any new guides, cards, or templates. Use it the morning of the workshop and any time you come back to the repo after a break.
The workshop ships two parallel tracks. 4D is the standard one-day path everyone runs (orient, connect, automate). 8D adds the deeper architecture (Cairns vault, Librarian, fleet patterns) for students who want to go all the way. Each card is badged with the track it belongs to.
The five-stage shape of the workshop library: 01 Get oriented, 02 Set up your machine, 03 Connect your world, 04 Run autonomously, 05 Capstone. Each card belongs to one stage, shown in its breadcrumb.
The deployed workflow at the end of the workshop: the Morning Brief. Everything you build during the day plugs into it, see Morning Brief.
Where to go next.
You have the words. The natural next moves: