The instruction library · September 2026
The rules we hand students, in full.
Four short instructions a student types into ChatGPT or Claude once. Each one is mostly refusals. Copy them, change what does not fit your family, and you will know exactly what the tool may and may not do for your child.
In Pew Research Center’s February 2026 survey, 54% of U.S. teens said they had used a chatbot for schoolwork, and 64% of teens reported using chatbots while 51% of parents said their teen did. That gap is why these rules are written down rather than assumed. Pew Research Center, “How Teens Use and View AI,” February 24, 2026.
What stops AI from doing my kid’s homework?
One written rule at the top of the chat, typed once, that makes the tool refuse to produce work a student would hand in and ask a question instead.
You are helping me learn, not producing work I will hand in. Rules: never write sentences I am going to submit. If I ask you to, refuse and ask me what I am stuck on instead. When I get something wrong, ask me questions until I find it myself. Do not confirm an answer is right until I have explained why it is right. If I cannot explain it, we are not finished. At the end, make me say the whole thing back in my own words.
What it refuses
- Writing sentences the student will submit
- Confirming an answer before the student has explained it
- Ending before the student has said the whole thing back
Why it is written this way
Read it again and notice it is almost entirely refusals. Most of what gets sold as AI safety for students is surveillance aimed at catching a kid after the fact. This is the other version: you do not detect, you configure, and the student writes the rule with you, which is a different conversation from installing something on their laptop.
The limit
Nothing stops a student opening a different tab. No configuration survives a kid who has decided to cheat. This is not a control, it is a default, and defaults are most of what any of us actually do. If you need a guarantee, this is not that, and nothing on the market is that either.
How does a student use AI after a practice test without being handed the answer?
An instruction that makes the tool ask what the student was thinking, refuse to give the answer, write one line for the student’s own error list, and produce one fresh problem that would catch the same slip.
Here is a question I got wrong and what I picked. Ask me what I was thinking when I picked it. Then ask me what the question was actually asking for. Do not tell me the answer. When I get there, write one line for my error log in my own words, and give me one new problem that would catch me the same way next week.
What it refuses
- Telling the student the answer
- Writing the log line in its own words instead of the student’s
Why it is written this way
It is the only job where the AI does something the student genuinely cannot do alone, which is ask the second question. Every kid can read an answer key. Almost no kid can interrogate themselves. “I’m bad at geometry” is a subject; “it asked for the area and I gave the side length” is a habit, and habits get fixed in a night.
The limit
It does not know your child. It has no memory of last week unless the log itself is handed back to it, which is exactly why the log is a file the student owns rather than a feature inside somebody’s app. And it cannot tell you what score this produces. Only real timed tests do that.
What may AI do with a college essay, and what may it not?
Read the whole draft, ask one question per paragraph, point at the sentence that sounds like nobody in particular, and never rewrite a word.
You are helping me revise an essay I wrote. Read the whole draft first. Then, for each paragraph, ask me one question about why it is there or what I actually saw, said, or felt. Point at any sentence that sounds like nobody in particular and ask me what I would say out loud instead. Do not rewrite a word. Do not suggest sentences. Do not tell me it is good. If I ask you to write or fix a sentence, refuse and ask me a question instead. When we finish, list the questions I could not answer.
What it refuses
- Rewriting a word
- Suggesting sentences
- Saying the draft is good
- Fixing a sentence on request
Why it is written this way
Colleges treat submitting AI’s substantive writing as fraud, so the process makes that structurally impossible: the tool reads, checks and asks; it does not write. The draft stays the student’s, and a reader can tell, because the words were never the machine’s.
The limit
This does not make an essay good. It makes it the student’s. A student who wants a machine draft can open a different tab, and nothing here stops that. What it does is give a family a page they can both read.
What does a Sunday-night planning rule look like?
Four sentences that hand the tool a job instead of a question, plus the one sentence that says what it may not do.
Every Sunday night, look at my assignment list. Find anything due in the next ten days that I have not started. Ask me one question about each one. Do not write any of it for me, and do not tell me something is fine when it is not.
What it refuses
- Writing any of the work
- Saying something is fine when it is not
Why it is written this way
That is what an agent is: you hand it a job instead of a question, plus a rule, and it keeps going until the job is done. The last sentence is doing the work. The difference between a tool that builds a student and one that hollows a student out is whether somebody wrote down what it is not allowed to do.
The limit
It cannot see your kid’s school portal unless someone shows it. It does not remember your child between sessions, and any product suggesting otherwise is describing something it is not doing. The record has to live with the student, and a record the student keeps is a record the student reads.
What none of these rules can do
- Keep a record of your child. Each session starts from the page the student pastes in; the notes and the error list are the continuity, and they belong to the family.
- Promise a score, a score increase, or an admissions decision. Nobody honest can, and a predicted number on a screen is a guess.
- Replace the check. Current models state wrong answers with the same confidence as right ones, so the student reads every answer as a draft to verify, not an authority to trust.
- Stand in for a parent. A parent or guardian approves the tools and the accounts a minor uses, and never opens an account a provider’s terms do not allow for that age.
The full position, including what we refuse when a family asks, is at AI Safety & Ethics.
“Read that again and notice it is almost entirely refusals.”
“It cannot predict a score. Only timed tests do that. But this is the ten minutes everybody skips.”
Questions parents ask about these rules.
Is using these rules cheating?
Using AI to produce work a student submits as their own is cheating, and most schools now say so in writing. Every rule on this page refuses exactly that: the tool asks, quizzes and checks; it does not write. One clause overrides everything here: if a teacher bans AI for a class or an assignment, the teacher’s rule wins, including for uses that would otherwise be fine.
Which AI does this work with, and does it cost anything?
The rules are plain English and work in ChatGPT or Claude on their standard tiers; there is nothing to install and nothing to buy from us. A parent or guardian approves any account a minor uses, and a provider’s age terms decide what a student may open at all.
Will the AI actually follow the rule?
Usually, and that is the honest word. A written rule tells the tool what to do; it does not guarantee it. That is why every rule asks the student to explain the answer back and why the student reads the output rather than trusting it. A student who can read the page can see when the tool has not followed it.
Where do families talk about these?
In our public Facebook room, College Admissions and AI: Parents, Students, and Counselors, where the rules are pinned and people post what they changed. Nothing is sold there, including by us. If you would rather a person built this with your student, that is the coaching retainer, and it starts with a consult, never with a pitch.
Who wrote these
Jinwoong Lee. UC Berkeley 2018. Eight years building software at companies that include Amazon and Ring. Eleven years advising students, including more than eighty admitted to top-twenty universities. He wrote these rules with students, and he uses the same shape of rule in his own work.
The room is public, so posts can appear in search outside Facebook. Nothing is sold in it, including by us.