Glossary

Agent. An AI that can take actions on your behalf — read your messages, send things, call other tools — not just answer questions. Always-loaded context. General rules and reference material the agent carries into every session. Keep it tight, because every word in there costs usage every session. Click moments. Three behavioral shifts that mark a user crossing from "I'm trying this" to "this runs my work": writing bigger prompts, making skills, asking the agent for ideas. Cheap model / smart model. Two tiers of language model. Cheap models are fast and inexpensive but less capable. Smart models reason better but cost more. Use cheap models for gathering and routine tasks; reserve smart models for judgment work. Get it done by any means necessary. Magic phrase. Used to push back when the agent says it can't do something. Works almost every time. Going forward. Phrasing that converts a one-off correction into a permanent rule. "Going forward, always X" gets saved; "actually, do Y this time" doesn't. Cron job. A scheduled task that runs at an exact time (e.g. 7 AM daily, Monday 9 AM, first of the month). Different from a heartbeat, which fires on an interval rather than a precise clock time. Triggered by saying "set a cron job for..." Hard guardrail. A restriction enforced by the platform or the connection itself, not by the agent. The agent literally cannot do the action (e.g. a read-only API scope means write actions are refused at the API). Use for anything irreversible or money-touching. Heartbeat. A scheduled task that wakes the agent up on an interval to check a list and decide whether anything needs your attention. Turns a reactive agent into a proactive one. Internal messages. Odd-looking messages the agent meant for itself or a sub-agent that occasionally leak into your view. Harmless visual noise. Magic phrases. Short phrases that solve the most common moments of friction. The core set: "get it done by any means necessary," "search deep," "going forward — keep that in always-loaded context," "make the required changes so you do this from now on," "do what you need to do so you never forget you have access to this tool," "make a skill for this," "update the skill with my feedback," "set a cron job for..." and "OK, go do it." Memory. Everything the agent remembers between sessions. You don't manage it directly — the agent saves what it needs and retrieves it when relevant.

Meta-prompt. Asking the agent to write a prompt for you. "Write me the absolute best prompt for X." Often beats writing the prompt yourself. Orchestrator. A configuration where the main agent never does work itself — it only spawns sub-agents and synthesizes their results. Keeps the main agent responsive and its context clean. Planning. A short conversation before executing a complex task — the agent asks clarifying questions, suggests angles, and gives a confidence score on the plan. When the confidence hits 90%+, you say "OK, go do it." See Chapter 4. Project. A named piece of work the agent is tracking — with its own plan and running log. Picks up where you left off across sessions. Read-only. A connection mode where the agent can see data but not modify it. The default for any new tool until trust is established. Search deep. Magic phrase used when the agent says it doesn't remember or can't find something. Tells it to look harder through everything it knows. Skill. A saved recipe for one specific task, stored as a plain-text instruction file in the agent's library. Loads on demand. The most important unit of customization. Soft guardrail. A rule the agent obeys because you told it to, not because the platform stops it. Lives in always-loaded context, a skill, or a prompt. Strong for workflow preferences; weaker than a hard guardrail for anything irreversible. Sub-agent. A temporary agent the main agent spawns to do a specific piece of work. Has its own context. Returns results to the main agent. System audit. A maintenance prompt run weekly. The agent compresses its own rules, moves procedural content into skills, cuts redundancy, and reports on usage. Prevents drift. Token. The unit of work the model charges for. Input tokens are what you send; output tokens are what it generates. Token discipline is the heart of cost control. Trigger word. A prefix (often @AssistantName ) that controls when the agent responds in group chats. Without the trigger, messages are stored as context but ignored. Plain-English definitions of every term used in the manual.

Closing note

You will not master this in a week. You will not master it in a month. After about three months of daily use, you will look up and notice that you are running half your business through it without thinking about it. That is the goal. Get there.

The tools will keep evolving. The models will keep improving. The conventions will shift. The principles in this manual are about your relationship with the agent — how you talk to it, how you teach it, how you keep it useful — not about any specific software. Those principles are durable. Use the agent every day. Correct it when it's wrong. Save what works as a rule or a skill. Ask it for ideas. Push back when it says no. Do that and you'll be fine.