Vocabulary

The piece of software that actually produces the words. It has read an enormous amount of text and learned to predict what comes next, so it can write, summarize, and answer in fluent language. When someone asks “which model are you using,” they mean this engine doing the thinking, not the chat window in front of you, and newer or larger models tend to be more capable but cost more per use.

What you type to tell the AI what you want. It can be a single question or a full set of instructions with background, rules, and the format you want back. The quality of what you get out is mostly set by what you put in: a vague prompt gets a vague answer, and a clear, specific prompt with the right context gets work you can actually use.

An AI that does not just answer a question but takes steps to finish a task for you: reading files, drafting documents, checking its own work, and asking when it is stuck. Think of it as the difference between a search engine and an assistant. An agent can carry a whole job from start to finish instead of leaving you to copy and paste between a dozen chat answers.

A repeatable job you hand to an AI agent, with a clear starting point, a set of steps, and a finished product at the end. The word “agentic” just means “done by an agent that takes its own steps.” This is the whole point of the club: each workflow, like the Fundraising Pipeline, is a recipe you can reuse, with the same inputs, the same steps, and a reliable result every time instead of starting from a blank page.

Everything the AI can “see” at once while it works: your instructions, the documents you gave it, and the conversation so far. The context window is the size limit on that working memory. If a task needs more material than fits in the window, the AI may lose track of earlier details, so when answers start drifting it often helps to start fresh or trim what you have pasted in.

The small chunks of text the model reads and writes in. A token is roughly a short word or part of a word, so a page of text is a few hundred tokens. Tokens are the unit you get billed in and the unit the context window is measured in, so “this is too many tokens” just means “this is more text than the model can take in or afford to process right now.”

When the AI states something that sounds confident and reasonable but is simply not true: a made-up statistic, a quote no one said, a citation to a study that does not exist. The model is built to sound fluent, not to be right, so it will rarely say “I am not sure” unless prompted to. In funder-facing work this is the single biggest risk, which is why you ground its claims in your own sources and check them.

The AI uses the same calm, assured tone whether it is right or wrong. A polished, certain-sounding answer is not evidence that the answer is correct. It is tempting to trust fluent writing, so treat tone as style, not proof, and verify anything that carries a number, a name, a date, or a commitment before it leaves your hands.

Breaking a big request into smaller, ordered steps instead of asking for everything at once. For example: first pull the funder’s rules, then check your project against them, then draft the narrative. AI does noticeably better on a chain of clear small steps than on one giant vague ask, so learning to break work down is the single highest-leverage skill for getting good results, and it is also how you stay in control of the output.

The bundle of your own materials you hand the agent to work from: the funder’s guidelines, your budget, past reports, program notes, whatever the task needs. It is the “here is everything you need to know” folder, and it is what keeps the AI honest. When the agent works from your documents instead of its own memory, the output reflects your real numbers and your real program, not a plausible invention.

Tying the AI’s answers to specific source material you provided, so each claim traces back to a real document rather than the model’s general knowledge. Grounded answers are checkable: instead of “trust me,” you get “this came from page 4 of your budget,” which is exactly what a careful reviewer or funder needs.

A citation is the pointer the AI gives back to where a claim came from. A trace, or trace table, is the full map showing line by line which source backs which statement in the finished document. A trace table turns “the AI wrote this” into “here is the evidence behind every line,” which is what makes an AI-assisted report defensible and lets a second person verify it fast.

Improving the result through rounds of feedback rather than expecting a perfect first draft. You read what it produced, tell it what to fix, and it revises. The first output is a starting point, not a final answer, and the people who get the most from AI treat it like a draft they shape, giving specific corrections like “shorten this section and use the 2025 figures.”

Asking the AI to return its answer in a fixed shape, like a table, a checklist, or labeled fields, instead of a loose paragraph. Structure makes the result usable and easy to scan: a budget as a clean table, or a review as a pass/fail checklist, drops straight into your work instead of needing to be reformatted by hand.

When the agent does more than write text: it opens a file, searches the web, runs a calculation, or saves a document. These extra abilities are its “tools.” Tool use is the difference between an AI that talks about your budget and one that actually opens the spreadsheet, checks the math, and writes the corrected version back out.

A common standard that lets an AI agent plug into outside tools and services, like your document storage or a database, the way a USB port lets any device connect to any computer. You do not need to understand the plumbing, but the name explains how an agent reaches your real files and systems: when a tool says it “supports MCP,” your agent can connect to it.

Keeping a person at the key decision points so the AI proposes and a human approves. The AI drafts, suggests, or flags; you decide what actually goes out or gets done. For anything with stakes, like money, commitments, or a funder relationship, you stay the final authority, and the AI speeds up the work without ever quietly making the call for you.

How much the agent is allowed to do on its own before checking with you. Low autonomy asks permission at every step; high autonomy runs the whole task and reports back at the end. You set this dial based on trust and stakes: a first run on an important grant deserves low autonomy and close watching, while a routine, well-tested task can run with more freedom once you have seen it work.

The deliberate step of checking the AI’s work against your sources and your judgment before you rely on it: confirming the numbers, the names, the claims, and the logic. Verification is the habit that makes AI safe to use for real work. The goal is not to trust the AI, it is to build a process, like a trace table and a quick review, that makes checking fast and routine.

The rules and limits you set so the agent stays inside safe boundaries: which files it may touch, what it must never do, and when it has to stop and ask. Good guardrails let you delegate with confidence. You decide up front, for example, “never send anything externally” or “always show me the budget before saving,” and the agent works within those lines.