The five levels of working with AI
Most people jump straight to automating their AI, and that's why it ends up a mess. There are five levels, and the order matters: ask, think with it, teach it, hand it a job, let it run. Each one rests on the one before. Find the level you're on, do both prompts for it, and stay there until it works by hand. Everything in [square brackets] is yours to swap out.
You'll have tried every level once, with your own material, and you'll know which one to work on next.
Before you start
Open a new chat in Claude or ChatGPT. Levels 1 to 3 work in either one, free or paid.
Levels 4 and 5 are different. They need a tool that can read files you give it or connect to an app like Gmail, and for level 5, run on a schedule. In Claude, scheduled tasks are on the paid plans (Pro, Max, Team and Enterprise), according to Anthropic's help page. For ChatGPT or any other tool, look at its settings or help pages to see what your plan allows before you start. If you can't do levels 4 and 5 yet, levels 1 to 3 are still worth your week.
Use real work for the examples, but leave out anything confidential: client names, account numbers, other people's private details.
Level 1 · Ask: change who it's for
You type, it answers. Everybody starts here, and that's fine. The first trick is that the answer changes a lot when you say who it's for. Pick one thing you half-understand from work, like a word in a contract or a number on a report, and send this three times in a row.
Level 1 · Ask · Same question, three audiencesExplain [the thing you're trying to understand] to a curious 10-year-old. (Then send it again, ending with "to someone who works in my field." Then once more, ending with "to someone who has never heard of it and is a little nervous.")
What to change: the part in brackets, and the three audiences. Use people you actually talk to.
Three answers to the same question, each with different words, examples and depth. Notice which one you understood fastest. That's the audience to name next time.
Level 1 · Ask: tell it the job and the shape
A good question names three things: who you want it to be, what you want back, and who it's for. Try all three at once on something you need this week.
Level 1 · Ask · Role, format, audienceAct as a patient [teacher / editor / accountant / recruiter]. I need [a list / a table / three short paragraphs / a checklist] about [your question]. It's for [me / my boss / a client who is not technical]. If anything in my question is unclear, tell me what's missing instead of guessing.
What to change: the role, the format and the reader. The last sentence stays; it keeps the AI from filling gaps with guesses.
An answer in the shape you asked for. If it's generic, add who it's for and what you'll do with it.
Level 2 · Think with it: find the holes
Now stop asking for answers and start asking for pushback. Use it as a sparring partner for a plan you already have: a project, a launch, a decision you keep going back and forth on. If you have an old chat about it, use that one.
Level 2 · Think with it · Find my blind spotsHere's my plan: [paste it, or describe it in a few sentences]. Find my blind spots and risky assumptions. Rank them from most to least likely to hurt me, and for each one, tell me how I could check whether it's true before I commit.
What to change: the plan. Keep the ranking and "how I could check", or you'll get a worry list with nothing to do about it.
A ranked list of what could go wrong. Pick one, just one, and check it this week. If a point doesn't make sense, ask "why do you think that?"
Level 2 · Think with it: question the task itself
The most useful thing it can tell you is that you don't need to do something at all. Before you hand any task to an AI, ask whether the task should exist.
Level 2 · Think with it · What can I drop?Here's what I do every week to get [the result you want]: [list the steps]. Which of these steps could I stop doing, combine, or do less often without hurting the result? Tell me what you'd cut first, and what I'd lose.
What to change: the result and the list of steps. Be honest about the boring steps; they're the ones worth asking about.
At least one step you could drop. Skipping a task saves more time than any automation, and it's the cheapest thing on this page.
Level 3 · Teach it: brief it like a new hire
Treat it like someone on their first day. You wouldn't say "write the weekly update" and walk away. You'd give them the goal, the background, an example of what good looks like, and your limits. Pick something you write or do often, like an email, a post or a summary.
Level 3 · Teach it · The new-hire briefI want you to [the task, for example: write my weekly client update]. Background: [who it's for, what they care about, what happened this week]. Here's an example of one I liked: [paste a past one, or describe what good looks like]. Limits: [length, tone, things never to say, things never to make up]. Before you start, ask me 5 questions about anything you'd need to get this right.
What to change: everything in brackets. The example matters most. Without one, it falls back on the average of everything it has read.
Five good questions before it writes a word. Answer them, read the draft, then tell it what to keep and what to change. Do two or three rounds. That feedback is the teaching.
Level 3 · Teach it: save what it learned
Once the draft is right, save the lessons so you never explain them twice. Say this in the same chat, before you close it.
Level 3 · Teach it · Write down the rulesBased on everything we just did, write me a short set of instructions I can paste at the start of a new chat so you do this the way I like it. Include the goal, who it's for, the style we settled on, the example, and the mistakes you made that I corrected.
What to change: nothing. If your tool has saved instructions, projects or skills, store them there; if not, keep them in a note and paste them in.
A short set of instructions in your own words. Open a new chat, paste it in, and ask for a new draft. If it misses something, tell it, and ask it to update the instructions.
Level 4 · Hand it a job: files in, report out
Now give it a real piece of work with more than one input. Pick a job you'd normally spend an hour on: a few documents, a spreadsheet and some notes, and one finished piece at the end. Add the files, then use the same brief as level 3.
Level 4 · Hand it a job · The reportI want you to read the files I've added and write [a one-page summary / a status report / a list of what changed since last time]. Background: [what these files are and what decision the report supports]. Example of the format I want: [paste or describe it]. Limits: Cite the file for every key finding. Say so when something is missing instead of guessing. Don't change my original files. Before you start, ask me 5 questions.
What to change: the deliverable, the background and the format. Keep the three limits; they're what make it checkable.
A finished draft that names its sources. You're still the one who signs it, so the next prompt is the important one.
Level 4 · Hand it a job: make it show its work
The AI sounds equally sure when it's right and when it's wrong, so you check, every time. Ask it to make that easy.
Level 4 · Hand it a job · Check the reportGo back over your report. List every number and every claim in a table: what it says, which file it came from, and how sure you are. Mark anything you inferred rather than read directly. Then tell me what you couldn't find.
What to change: nothing. Then open two or three of the sources yourself and compare.
A table you can check in ten minutes. Anything marked as inferred is a guess until you've confirmed it. Fix what's wrong, then save the instructions the same way you did in level 3.
Level 5 · Let it run: connect one app, read-only
Only start here once level 4 has worked by hand a few times. First, connect one app you actually use, like Gmail. In Claude these are called connectors; in ChatGPT, apps or plugins. When you connect it, turn off anything that can send, change or delete, while you're learning.
Level 5 · Let it run · Look, don't touchRead the 3 most recent items in [Gmail / your calendar / the app you connected]. Show me the date and who each one is from. Don't reply, send, move or change anything.
What to change: the app. This only proves the connection works.
Three real items from your own account, read back to you. If it can't see them, the connection or the permission is the problem, not your prompt.
Level 5 · Let it run: put it on a schedule
Now take the job you've done by hand and ask your tool to run it on its own, with the instructions you saved. The schedule only decides when it starts; the checking is still yours. Your first few runs should end in a draft for you to read, not an action taken for you.
Level 5 · Let it run · The scheduleCreate a scheduled task that runs the instructions we saved for [the job] every [Monday] at [8:00 AM] in [my time zone]. Have it prepare a draft for me to review, and don't send or change anything on its own.
What to change: the job, the day and the time. If your tool can't schedule, it should say so; some only run while your computer is on and awake.
A task on your calendar with a next run date. Use "run now" once to test it, and read the first few results closely before you stop watching.
Automate last, not first. Skip to level 5 and all you get is mistakes, faster, on a timer. Five-levels framework from Sabrina Ramonov (@sabrina_ramonov). The prompts and this page are mine.
If you get stuck
Which level am I on?
Level 1: you ask and it answers. Level 2: you test your ideas against it. Level 3: it knows how you like things done. Level 4: it does a whole job and you check it. Level 5: it does the job on a schedule. Whichever you can't do yet, that's the next one.
The answers are generic.
Say who it's for, what you'll use it for, and paste in one example of what you'd call good.
It made something up.
Reply: "Which part of that did you not find in what I gave you? Take it out, and ask me instead." At level 4, send the check-the-report prompt every time.
I can't connect an app or schedule a task.
That's usually a plan or a permission setting, not something you did wrong. Stay at level 4 and run it by hand each week. That alone is a real win.
Does it actually learn?
Not the way a person does. The saved instructions are a note it reads next time. When you correct it, ask it to update the note, then test the note in a new chat.
Your final check
Tick these as you go. The ticks reset when you leave this page.
Keep going when you’re ready.
Pick another guide for the next problem you’re working on.