AI Level Guide

Cheat Sheets

New to any of this? Start here. Five sheets, front to back: what these tools actually do and which one fits the job, what never goes into them, how much thinking room to give the model, what your team means by the words they use for saving and sharing work, and that same loop watched happening one step at a time. Nothing here assumes you have used GitHub, or written code.

Start here if AI tools just landed on your desk and nobody explained them. This is what the thing actually is, the handful of shapes your work with it will take, and where to point it first. No install, no jargon, and nothing here needs a technical background.

What it is

  • It reads anything you give it. Paste a forwarded thread, drop in a spreadsheet, hand it a sixty-page PDF. It will read all of it and answer questions about what is in there.
  • It writes a solid first draft. Emails, summaries, specs, updates, agendas. A draft you edit is a different starting line from a blank page, and that difference is most of the time saved.
  • It thinks through messy problems. Half-formed, ambiguous, no obvious next step. Describe the mess the way you would to a colleague and work it out in the open.
  • It does the tedious parts. Reformatting, cross-checking two lists, pulling twenty action items out of an hour of notes. The work that is not hard, only long.

What it is not

  • Not a search engine. It is not looking things up on your behalf by default. Some versions can search when you ask them to, but what it is genuinely strong at is reasoning over what you hand it.
  • Not connected by default. It does not know your inbox, your drive, your calendar, or your tracker unless somebody deliberately connects it. Until then it knows only what you paste or upload.
  • Not always right. Treat the output as a strong draft rather than an answer. You are still the reviewer, and your name is still the one on the thing you send.
  • Not a place for raw sensitive data. Mask identifying detail before it goes in. That is the whole of tab 02, and it is a two-minute read.

Three ways you will use it

  1. 01Just ask

    Type it or say it out loud. Paste in the messy thing — the forwarded thread, the half-written paragraph, the notes you took too fast — and ask for what you want out of it. No setup, no format to learn, no right way to phrase it.

  2. 02Bring your files

    Upload the spreadsheet, the PDF, the deck. It reads them, works on them, and hands finished files back — not a description of a deck, the deck.

  3. 03Connect your tools

    Once an admin switches on a connector, it works directly in the systems your work already lives in. Nothing gets pasted in and nothing gets downloaded out.

Almost everything else you will read about is one of these three wearing a different name. There is far less to learn here than the vocabulary suggests — the next section is the one that matters, and it is four rows long.

Match the tool to the job

Four shapes. The last column is the one to read — you arrive with a task, not with a shape in mind.

  • Chat

    Reach for it when

    A one-off question, a first draft, or something you want to think out loud about.

    What you end up with

    An answer, in the window, now.

    The tell

    You will not need this again tomorrow.

  • Skill

    Reach for it when

    A task you do over and over, the same way each time.

    What you end up with

    A saved playbook in plain English that you — or a teammate — can point the AI at.

    The tell

    You have explained the same process twice.

  • Scheduled task

    Reach for it when

    Something that should happen without you, on a cadence.

    What you end up with

    A standing job. The Monday rollup is already waiting when you open your laptop.

    The tell

    You keep remembering it late.

  • Connector

    Reach for it when

    The answer lives across several systems, or the result has to land back in one of them.

    What you end up with

    Work that reads and writes where it already lives, with no copy-paste in the middle.

    The tell

    You are the integration.

Most people meet all four as chat, and then wonder why AI never quite sticks. The move from personal use to something a team leans on is mostly this: noticing which of the four a piece of work actually wants. If you find yourself running the same chat every Monday morning, that is not a chat — it is a scheduled task you have been performing by hand.

Where the work happens

Same model, three different rooms. This is a tool choice, not a ranking.

  • Cloud / chat

    A browser tab or a desktop app. You talk, it answers. Your files go up to it and come back finished.

    Good for

    Questions, drafts, and analysis of something you can attach. Anything you would finish in one sitting.

    When you would reach for it

    Where everyone starts, and where a great deal of good work stays for good. L0 and L1 live here almost entirely.

  • Local / Cowork and Code

    It works on real folders on your own machine, several steps at a time, showing you what it is doing as it goes. You steer, and you can stop it.

    Good for

    Multi-step jobs across many files: reorganising a folder of documents, producing a set of deliverables, following a process you wrote down once.

    When you would reach for it

    Worth opening the first time one sitting in a chat window stops being enough. Common from L2, but nothing stops an L1 trying it — the learning curve here is the file system, not code.

  • Terminal / IDE

    It lives where code lives, alongside the same files an engineer has open. Spare and text-first, with none of the chat furniture.

    Good for

    Work that touches a codebase, and anything where the version history is part of the job — which is what tab 04 is about.

    When you would reach for it

    Reach for it when the work genuinely is in code, or when you are pairing with someone whose work is. Common from L3, and it is a tool choice rather than a rank.

None of these is the advanced version of the other two. They are three different rooms with the same model in each. Pick the one that matches where the work actually sits: in your head, in a folder, or in a codebase.

Everyday wins

Easiest first. The top one needs no setup, no permission, and no judgment call — you were already in the meeting.

  • Meeting minutes and action items

    The easiest win there is, and the one to do first. Hand over a transcript or your own scrappy notes and ask for decisions, owners, and open questions as three separate lists.

  • Status rollups

    Several workstreams, one update, every week. Give it the updates in the state they arrived — chat messages, forwarded mail, a half-filled tracker — and ask for the version your manager will actually read.

  • A risk and issue log out of a messy thread

    A long, argumentative email thread usually contains a risk register nobody wrote down. Ask for it as a table: the risk, who it lands on, how likely it is, and what would reduce it.

  • Real files, not descriptions of files

    It can hand back a PowerPoint you can present, an Excel workbook whose formulas still calculate, a Word document, a PDF. You get the artifact, not a summary of what the artifact would contain.

  • Artifacts

    Describe a small tool and get a working, clickable version of it in a couple of minutes — a calculator, a checklist, a way to look at some numbers. Faster than three paragraphs explaining what you mean.

  • Data analysis without formulas

    Upload the file, ask your question in English, get charts and the reasoning behind them. Read the reasoning: it tells you what it assumed, and the assumptions are where you catch it being wrong.

Make it yours

  • Projects (Folders)

    A workspace that remembers. Give it its own files and its own standing instructions — how you write, what your readers already know, what to never do — and every conversation inside it starts from there instead of from nothing. Share the project and a whole team gets the same shape of output without anyone having to remember the instructions.

  • Skills (A text file of instruction)

    If you have explained the same process twice, write it down once. A skill is that write-up: plain English, no code, saved where the AI reads it. Next time the task comes round you name the skill instead of re-explaining the job.

  • Connectors (talk to apps)

    A connector lets the AI read and act in the systems your work lives in, rather than in a copy you pasted. Three things people reliably get wrong about them:

    • It acts as you, and creates no new access. Anything you cannot open, it cannot open either.
    • An admin turns them on. Ask rather than assume — the answer is often yes, and asking takes an afternoon rather than a quarter.
    • Anything that writes or posts asks for your approval first. Reading is quiet; sending, posting, and changing come with a confirmation step.

  • Voice

    The fastest way to get something out of your head. Talk for two minutes about a problem the way you would to a colleague, then ask for it back with structure. Rambling and half-sentences are fine — you are not required to tidy your thinking up front.

  • Sharing

    You can turn a conversation into a link and send it. Worth knowing exactly what travels with that link: everything in the chat.

    The file you uploaded near the top goes with it. So does the tangent in the middle where you were thinking out loud, and the draft you abandoned. If you know you are going to share, start a clean chat and do the thing once, properly.

The words differ by vendor; the shapes do not. A project is a Copilot agent or a Gemini gem. A connector is Copilot wired to SharePoint, Outlook, and Teams, or Gemini inside Google Workspace. If your organisation standardised on one of those, read this tab in its vocabulary — none of these four is Claude-only.

How to ask well

  • Say who it is for. The same summary written for a customer, an engineer, and your manager's manager is three different documents. Name the reader and you get the right one first time.
  • Give it the context you would give a new hire. Not a briefing pack — a paragraph. What the project is, what has already been decided, and what you are not allowed to change.
  • Say what good looks like. Length, format, tone, and one example of a version you liked. “Like this, but for next quarter” is a complete instruction.
  • Ask for the options before the answer, when the problem is fuzzy. For anything unclear, ask for two or three approaches and their trade-offs, then pick one. Tab 03 is the longer version of this idea.
  • Push back. The first answer is a draft. “Too generic, and you missed the budget constraint” gets you a better second one faster than starting over.
  • Ask it what it needs. “What do you need from me to do this well?” is the highest-yield question on this page when you do not know how to start.

Pick one, do it this week

One is enough. Pick whichever matches something already on your plate this week, and do it on real work rather than on a test case — a practice run teaches you almost nothing.

  • Take the notes from your last meeting and ask for decisions, owners, and open questions.
  • Upload the spreadsheet you keep re-explaining, and ask what it notices before you tell it what to look for.
  • Write your standing instructions — who you are, who your readers are, how you want responses — and save them in a project.
  • Take the process you talked a colleague through last month and have it written up as a skill.
  • Find the report you rebuild every Monday and ask whether it could be a scheduled task.