Before You Start: A Clear Map of AI

Ten minutes to understand what AI is, the words that matter, and which tool fits which job.

Framework Free · 10-minute read · Updated Nov 1, 2026

Everyone is talking about AI, and the advice comes from every direction. Before you try another tool or copy another prompt, spend ten minutes here. This page gives you the map: what AI is, the handful of words that keep coming up, and how to tell which tool fits which job. Every tutorial on this site builds on it.

The map answers five questions:

1. What can AI do well, and what can’t it?
2. What do words like prompt, context, connector, Agent and Skill actually mean?
3. Which level of AI am I using, and what am I really doing each time I use it?
4. Which tool, from which company, should I use for which job?
5. What should I never hand over to AI?

What AI is (and isn’t)

Think of AI as an intern on their first day: they’ve read a huge amount, they work incredibly fast, and they know nothing about your situation. They’re great with words. They’re not guaranteed to be right.

Under the hood, it works like this: given the text you provide, it produces the response it thinks fits best. That one sentence explains both its strengths and its weaknesses.

One line to remember: AI gives you a first draft, not an answer.

Five basic concepts

The news and product pages throw a lot of jargon at you, but only five words really matter. In intern terms, they cover how you give instructions, what material you hand over, what access you grant, how big a job you let them run with, and what handbook you give them.

Put together: you instruct with a prompt, supply material and access through context and connectors, teach your standard way of working with a Skill, and let an Agent run the whole job. The first three are everyday basics. Agents and Skills are the next step, and each gets its own section below.

Three levels

Everything sold as “AI” today falls into three levels. The difference is how much the AI can do on its own. Once you know which level you’re using, you know what to expect, and what to check.

The band across the bottom is the AI many office workers actually meet every day: the assistant built into Word, Excel, Outlook, Teams or Google Docs. It isn’t a fourth level. It’s the same abilities through a different door. Rewriting a paragraph in Word is Level 1. Asking it to read a pile of emails is Level 2. Having it work through a task step by step is Level 3. If your company only allows this one, that’s completely fine: everything on this site still applies.

Most people start at Level 1. The first tutorials on this site focus on Levels 1 and 2. Level 3 comes once you’re comfortable with the loop and the boundaries.

The loop

Every time you use AI, you’re running the same loop: give context, give instructions, get a draft, check it, and go round again if it isn’t right. Most people only do the middle two steps, which is why their results are so-so.

When you check, ask yourself three questions:

1. Are the facts and numbers right?
2. Is anything important missing?
3. Is the tone right for the person reading it?

This loop is the foundation for the next two sections: an Agent is AI running this loop by itself, and a Skill is the version of the loop you’ve already perfected, saved for next time.

Agents: hand off the whole job

An Agent is what you get when you stop supervising the intern step by step and give them a goal instead. They work out the steps, find the material, use the tools, and hand you the finished result. It isn’t a different kind of AI. It’s the same intern, trusted with more.

How it differs from a normal chat

In terms of the loop: an Agent runs many rounds of “instruct, draft, check” on its own, and leaves only the final sign-off to you.

When to use an Agent

It’s worth handing a job to an Agent when all of these are true:

1. The job has several steps, or spans several tools: research, organize, build a table, write a document.
2. You know what the finished result looks like, and can describe it in a sentence or two.
3. The steps in between don’t need your judgment again and again.
4. The result is easy to check: you can tell at a glance whether it’s right.

When not to: anything done in one step (a chat is faster); work that needs your judgment at every turn; actions that are costly and hard to undo, like emailing a client, deleting files or paying for something. For those, at the very least, make it ask you first.

How to brief an Agent well

Briefing an Agent is like giving the intern a big assignment. Cover four things: the goal, the deliverable, the limits, and when to come back to you.

Goal: Help me understand how 3 main competitors changed their products and pricing in the last 6 months
Deliverable: A comparison table + a one-page summary, with a source link for every point
Limits: Use public web pages only; don't use my email or internal files
Ask me: If the scope is unclear, ask first instead of guessing

When it delivers, check it with the three questions from the loop. The first few times, practice on tasks and files that don’t matter.

Where you can use one, and what to watch for

All the major companies have built Agents into their products:

• ChatGPT: ChatGPT Work (launched July 2026) is good at working across cloud apps and can keep going for hours.
• Claude: good at working with files in folders you choose on your computer, and also runs on the web and mobile.
• Microsoft 365 Copilot: Agent Mode in Word, Excel and PowerPoint makes multi-step edits right inside your files.
• Gemini: Google announced a general-purpose work Agent in October 2026. For now only some companies have it; it will roll out to Workspace business customers.

Two things to watch for:

• It makes mistakes, and makes them confidently. Public tests have caught Agents miscounting numbers and deleting files. Letting go doesn’t mean not checking.
• Give it only the access the job needs. An intern doing competitor research doesn’t need your email password.

Skills: a handbook for your AI

A Skill is a handbook you write for the intern: “here’s how we do this kind of task,” written down once, which the AI opens whenever that task comes up. If you notice yourself typing the same context and instructions into the loop over and over, it’s time to turn them into a Skill.

How it differs from other ways of instructing AI

Because the AI only opens a handbook when it needs it, you can keep plenty of them on the shelf without overloading your intern.

When to use a Skill

Any one of these signs means it’s worth writing one:

1. You do the same task repeatedly, and explain it from scratch every time.
2. There’s a set format or standard: a template, brand rules, a checklist.
3. Several people on your team need the same result, and one handbook can serve everyone.

When not to: one-off questions, and anything you haven’t figured out yourself yet. The order is always the same: get it working by hand in a conversation a few times, then turn what worked into a handbook. Nobody writes a full manual for an intern on day one.

How to write a good Skill

A Skill is a folder, and the most important thing in it is one instruction file. It starts with a name and a line saying when to use it, which is how the AI decides whether to open it. After that come the steps and standards, plus any templates.

---
name: weekly-report
description: Use when writing the team's weekly report
---
1. Start with this week's 3 most important results, one sentence each
2. Then list risks and where we need support
3. Keep it under 300 words, and skip the self-praise

You don’t have to write it from scratch. In tools that support Skills, once a task has gone well, just say: “Turn what we just did into a Skill.”

Where you can use one, and what to watch for

Skills are a Level 2 and Level 3 ability. Anthropic introduced them in October 2025 and published them as an open standard that December. Besides Claude, tools such as ChatGPT, GitHub Copilot, Cursor and Gemini CLI now support them. Names and support differ between products, so check your tool before you start.

• Only use Skills from sources you trust. The intern does whatever the handbook says. In February 2026, security researchers found 341 malicious Skills on community sites. Read a shared Skill before you install it.
• Keep them up to date. When your template or process changes, update the handbook, or the intern will diligently keep doing it the old way.

What office workers use AI for

Several large surveys from the past year agree: office workers use AI most for four things. Looking things up, reading and summarizing documents, drafting and editing text, and analyzing information or helping with decisions. Handing over whole jobs is growing too: in a July 2026 US survey, about one in five workers said AI now does at least one task for them entirely.

The first seven rows below are the most common uses and the best place to start. Then come other everyday uses, and the last three rows are Agent territory.

What not to hand over yet: final decisions, numbers that must be exactly right, and situations that need a feel for people. The intern can help you prepare, but they don’t get to make the call.

Survey sources: ZipRecruiter 2026 AI Worker Report (Sept 2026, 1,500+ US workers), Microsoft Work Trend Index 2026 (20,000 AI users in 10 countries), Epoch AI / Ipsos (July 2026, 1,106 employed US adults), OpenAI study of ChatGPT use (Sept 2025).

The AI industry in one picture

The companies in the headlines fall into four layers. Picture a restaurant chain: you order at the restaurant, a chef cooks in the back, the chef rents a kitchen from a landlord, and the kitchen equipment comes from yet another company.

For office workers, three things are worth remembering:

1. You almost only deal with the top layer. The same chef often cooks for several restaurants: Microsoft’s Copilot, for example, uses OpenAI’s models. But different kitchen equipment makes different food. The layer wrapped around a model, which decides what tools it can use and how it works step by step, is technically called a harness. That’s why the same model can feel different in Copilot and in ChatGPT.
2. Don’t chase version numbers. The chefs upgrade every few months, just like a new batch of interns every season. Where they studied matters less than whether you know how to work with them.
3. Which restaurant your company lets you into matters more than which is best. The next section turns this into practical choices.

This picture reflects October 2026 and is updated every quarter. Sources: Meta’s Muse models (Axios, Oct 2026), OpenAI’s ChatGPT Work (Let’s Data Science, Jul 2026).

Which tool for which job

Pick a tool in this order: first, what your company allows; second, where your files and work live; only then, which one is “best”. The gaps between the big names are shrinking. Getting really comfortable with one matters more than chasing the newest. Switch interns every week and you’re always training a newcomer.

This table reflects October 2026 and is updated every quarter, like the industry picture. The cloud-apps-versus-local-files split for Agents comes from a third-party comparison, not from the companies themselves. Source: The 7 best AI agents in 2026 (Zapier, updated Oct 2026), ChatGPT Work vs Claude (o-mega, 2026).

Boundaries

Before you use AI, remember four boundaries: what you must not paste in, what you must check, what your company allows, and what’s ultimately on you.

1. What not to paste in: client personal data, unpublished financial figures, contract text, passwords and logins. A simple test: if this text appearing on a website outside your company would cause trouble, don’t paste it into a personal AI tool.
2. What to check: numbers, dates, names, quoted sources, and anything that will leave your desk. Even when the AI says it looked things up in your files before answering (a technique called RAG), it can still miss things or misread them. Ask where each point came from, and spot-check one or two.
3. What your company allows: find out which tools you’re allowed to use. Many companies only allow an enterprise version or the AI built into their office software, and enterprise and personal versions treat your data differently.
4. What’s on you: whatever the AI writes goes out under your name. The responsibility is yours.

Glossary

One plain-English line per word. Come back here whenever a term trips you up.

Sources

• Agents: Google’s Gemini agent for work (Reworked, Oct 2026), Copilot Agent Mode (TechSpot), ChatGPT Work vs Claude (o-mega, 2026)
• Skills: AI Agent Skills (SumProduct), Agent Skills: the open standard (inference.sh)
• Surveys, industry and tools: linked in each section above
• Inspired in part by two Chinese-language learning roadmaps: awesome-agentic-ai-zh and Hello-Agents (Datawhale). The text on this page is original.