Appendix
Glossary
The words, in plain English. Skim the first section now. Leave the later section until a word comes up.
Tagged later = from "Words you might hear later".
A
- Agent
- An AI you chat with that can also do useful work when you give it access and clear instructions. For this guide, think of it as a careful assistant who still needs checking.
- Agent harness or multi-agent systemlater
- A setup where several agents work together, with one leading and others doing specific pieces. Useful for bigger builds, not needed for the everyday wins.
- Automation toollater
- A service that connects several apps and runs a recipe. Useful later when work needs to hop across systems or run on a schedule.
B
- Brief
- The clear instructions you give before work starts: what you want, who it's for, what good looks like, and what must not happen.
C
- Coding agentlater
- A specialist AI tool that can build or edit software. Handy later for custom tools, not required for this guide.
- Compaction
- When a long chat gets too big, many tools shrink the older part down to a summary. It's useful, and details can fall out, so check anything important is still there.
- Connector
- A safe link between your AI tool and another account, such as email, calendar or files. Start read-only, check permissions, and remove anything you don't use.
- Context
- Everything the AI can see right now: your message, the chat so far, files, pasted notes and connected sources. Better context usually means better output.
- Context driftlater
- When a long chat slowly gets worse because old instructions and side paths pile up. Starting clean is usually faster than arguing with it.
- Context windowlater
- How much the AI can hold in one chat before older material starts to fall away. If it forgets or drifts, start a fresh session with a handover note.
D
- Dashboard or cockpitlater
- One calm page that holds your priorities, live work, blockers and parked items. Build it manually first, then connect read-only, then schedule a brief if it earns trust.
- Draft, confirm, execute
- The core safety habit. The AI drafts, you check, then you choose whether anything is sent, saved, booked or changed.
E
- Eval
- A check on an AI's work. A list of real tasks with what good looks like, which you rerun after every change. Test your setup shows you how.
G
- Guardrail
- A rule that keeps the AI inside safe lines, such as "never send without showing me first" or "do not invent prices".
H
- Hallucination
- The AI improvising a confident answer that is wrong. Check names, numbers, dates and facts before anything leaves your hands.
- Handover notelater
- A short summary the AI writes so you can move a job into a fresh session without losing the thread.
- Harnesslater
- The app an agent runs in. Different ones suit different jobs, and good habits carry across all of them.
- Hooklater
- A small script a tool runs before or after an action. Unlike an instruction, it doesn't depend on the AI choosing to follow it.
K
- Kit
- Every setup and template from this guide, on GitHub.
L
- Large language modellater
- The kind of trained system behind most chat tools. You don't need the theory to use it well.
- Least access
- The smallest permission set that still gets the job done. If read-only is enough, don't grant write access.
- Local modellater
- An AI that runs on your own machine instead of a hosted service. Useful for some private work, but often less convenient.
M
- MCP
- The Model Context Protocol, a shared standard for connecting AI tools to other apps. Think of it as a common plug.
- Memory
- Saved facts and preferences the AI can use in later chats, such as your tone, business name or preferred format.
- Modellater
- The engine doing the thinking. Different models have different strengths, costs and speeds.
O
- Orchestratorlater
- A lead agent that breaks a bigger job into pieces and pulls results together.
P
- Permissionslater
- The exact abilities you grant when connecting a tool, such as read email or create calendar event. Read the list before approving.
- Prompt
- What you type to the AI. A clear prompt gets a clearer answer.
- Prompt injection
- A malicious or accidental instruction inside content the AI is reading, such as an email that tries to make it leak data. This is why connected work stays draft-confirm-execute.
R
- Read-only vs read and write
- Read-only lets the AI look but not change. Read and write lets it act. Start read-only unless you have a strong reason not to.
- Repolater
- Short for repository, a folder of files with its full history, usually code.
- Routine
- A saved prompt with a clock on it, which runs without you asking. Also called a scheduled task. Routines shows you how.
S
- Session
- One continuous chat. If a session gets messy or long, ask for a handover note and start fresh.
- Skill
- A saved repeatable workflow. It can be a note you paste in, or a folder with a SKILL.md that your agent loads when the job calls for it.
- Standing instructionslater
- Rules the AI reads before every chat, such as your tone, format and safety defaults.
- Subagent
- A helper agent given one part of a job by a lead agent. It uses far more of your allowance than you'd think. More than one agent explains when one's worth it.
T
- Tokenlater
- The small pieces of text an AI reads and writes. Plans and bills are counted in them.
- Triage
- Sorting a pile into buckets so you know what needs action, what is information only and what can wait.
W
- Workplace-controlled AI setuplater
- A company-managed setup where sensitive data stays inside approved systems. If you handle regulated or confidential work, ask for this path.