# Learning with AI without losing the learning

The same kind of tool more than doubled learning gains in one trial and cut exam scores in another. Four setup rules of ours and a checklist, built from those trials, aim for the first.

Guides · v1.1 · draft · https://hayzlabs.com/guides/learning-without-losing-the-learning/

In short:
- Try each problem first, then paste our one-hint-at-a-time tutor prompt so the assistant gives hints, questions or checks, not answers.
- Attach your own correct material, such as notes or a textbook chapter, and keep sessions to about 25 to 50 minutes.
- End each session by answering questions from memory with the assistant closed, and schedule a review.

## Short answer

The same kind of tool more than doubled learning gains in one trial and cut exam scores by 17% in another, where a version that gave hints and not answers avoided the drop. We suggest four rules: try each problem before you ask, give the assistant your own correct material, keep sessions short, and test yourself with the assistant closed. A quick test: if you cannot say what you could do afterwards without the assistant, you probably used it to produce and not to learn.

**Who this is for.** Anyone learning a subject or a skill with an assistant, for a course or for work. No setup is needed beyond your notes or a textbook chapter.

## The same kind of tool can double learning or remove it

High-school students with unrestricted ChatGPT did 48% better in practice but 17% worse on the exam once it was taken away. A tutor version that gave hints, not answers, avoided the drop (about 1,000 students, Turkey; [Bastani et al., PNAS 2025](https://ideas.repec.org/a/nas/journl/v122y2025pe2422633122.html), randomised trial).

In a Harvard physics trial with 194 students, median learning gains with a carefully designed AI tutor were more than double those from an active-learning class, in less time (median 49 minutes). Only immediate learning was measured, on foundational material ([Kestin et al., Scientific Reports 2025](https://www.nature.com/articles/s41598-025-97652-6), randomised trial).

Students who used ChatGPT improved their essays most but learned no more than students with other help (117 students; [Fan et al., BJET](https://arxiv.org/abs/2412.09315v1), randomised trial). A better essay and more learning are separate outcomes.

Reading a good explanation can feel like learning. A test with the assistant closed is what shows whether it was.

## Four rules

> [!OUR RULE]
> 1. **Try first.** Attempt each problem before you ask for a method, and ask for a hint, a question or a check, not the answer.
> 2. **Give it correct material.** Attach your textbook chapter or your own notes, and ask it to stay within them.
> 3. **Keep sessions short.** About 25 to 50 minutes at a time.
> 4. **Test with the assistant closed.** Finish each session by answering questions from memory.
>
> The four rules are ours, built from the trials. No study tests them together.

The group that got hints and no answers avoided the exam drop. The Harvard result was for a carefully designed tutor, which we read as a reason to give the assistant good material to work from. Rule 4 comes from a review of ten common study techniques, in which self-testing and spacing practice out had the strongest evidence ([Dunlosky et al., 2013](https://www.psychologicalscience.org/news/releases/which-study-strategies-make-the-grade.html), review).

## Use a study mode, and use it as the rules say

Built-in study modes help with rule 1. Names and menus change.

| Product | Study mode |
|---|---|
| ChatGPT | [Study Mode](https://openai.com/index/chatgpt-study-mode/) |
| Gemini | [Guided Learning](https://blog.google/outreach-initiatives/education/guided-learning/) |
| Claude | Learning ([press report](https://engadget.com/ai/anthropic-brings-claudes-learning-mode-to-regular-users-and-devs-170018471.html)) |

In Claude.ai, check Anthropic's help pages for where the Learning option currently sits. In Claude Code (Anthropic's coding agent), the Learning output style "leaves small pieces of code for you to write yourself", marked TODO(human) ([Claude Code docs](https://code.claude.com/docs/en/output-styles)).

A plain prompt that asks for one hint at a time does the same job, in any assistant.

## The one-hint-at-a-time prompt

Attach your notes or chapter, paste this as the first message, then paste your first attempt (our suggestion; we have not tested it across assistants). Setup takes about 2 minutes.

```text title="One hint at a time · v1.0"
You are my tutor for <TOPIC>. I want to learn to do this myself.
Use only the attached material: <MATERIAL>. If it does not cover
something, say so instead of guessing.

Rules:
- Never give me the full answer or the full method, even if I ask.
- I will show you my attempt first. If I have not, ask me to try.
- Give one hint at a time: the smallest nudge that gets me moving.
  Then stop and wait for my next attempt.
- If my attempt is right, say so and ask me why it works.
- If it is wrong, point to the step that went wrong, not the fix.
- If I am still stuck after three hints, show me a similar worked
  example with different numbers, not this one.
- When I say "done", ask me <NUMBER> questions on what we covered,
  one at a time, for me to answer later with this chat closed.
```

Replace the following:
- `<TOPIC>`: what you are learning, such as `reading a balance sheet`.
- `<MATERIAL>`: the file you attached, such as `chapter 3 of my accounting textbook`.
- `<NUMBER>`: how many end-of-session questions you want, such as `5`.

If it hands you an answer anyway, reply "Hint only, please" and paste the prompt again in a new chat.

## A study session (illustrative)

Priya is learning to read a balance sheet for work. She sets a 40-minute timer, attaches the chapter and pastes the prompt with `<TOPIC>` set to `reading a balance sheet` and `<NUMBER>` set to `5`.

1. **Try first.** She works the chapter's first example on paper and writes: "Current assets are 120, current liabilities are 150, so the firm is fine." She pastes that.
2. **Hint 1.** The assistant replies: "Look again at which is bigger. What does the chapter say it means when current liabilities are larger than current assets?" She rereads the page and revises her answer: the firm may struggle to pay bills due within a year.
3. **Check.** It says she is right and asks why the comparison uses current items only. She answers in her own words, and it points her to the paragraph that confirms it.
4. **Questions.** At 40 minutes she types "done". It gives five questions. She copies them into a note and closes the chat.
5. **Test closed.** She answers from memory, scores 4 of 5, writes `4 of 5` on the checklist below and sets a review for two weeks later.

What came out: one revised answer she reached herself, one gap (the question she missed) to start the next session with, and no answer she did not work for.

## The checklist

Paste it into a note and fill it in at the start and end of a session:

```text title="Am I still learning? · v1.0"
Before:  I wrote what I will be able to do, and how I will test it
         without AI.
During:  I attempted every problem before I saw a method.
During:  The AI explained. It did not produce the thing I am meant to
         learn to produce.
During:  My source material is attached.
After:   I answered questions with the AI closed. Score: <SCORE>
Later:   I scheduled a review for <REVIEW_DAY>.
```

Replace the following:
- `<SCORE>`: your result, such as `4 of 5`.
- `<REVIEW_DAY>`: the day you will test yourself again, such as `2026-10-24`.

The "Before" line works best as one concrete sentence, such as "I can read a balance sheet and say whether the firm can pay its bills", tested with questions you wrote yourself.

## How to tell it worked

- You can do a problem of the same kind on paper, with the assistant closed.
- Your end-of-session score is written down, and a review day is in your calendar.
- Every answer in the chat came from your own attempt; the assistant gave hints, questions or checks.
- It has failed if you copied a method or answer from the chat. Start a new chat with the prompt.

## Limits

> [!LIMIT]
> The trials are short, most are of students in courses, and they differ in design: one measured only immediate learning. Whether gains last is a separate question the trials here do not answer. Our sources include no study of adults teaching themselves with an assistant, so the rules are our reading of the school and university trials. For a skill where a mistake could hurt someone, check what you learned with a qualified person. If the point of a task is the result and not the learning, the rules do not apply: see [When not to use AI](https://hayzlabs.com/guides/when-not-to-use-ai/).

## Try this next

1. Next time you start something you want to learn from, write "output" or "skill" at the top. (1 minute)
2. Run one session with the four rules and the one-hint-at-a-time prompt, and fill in the checklist. (45 minutes)

## Sources

1. Kestin et al. Scientific Reports, 2025. Randomised trial, 194 students. [Link](https://www.nature.com/articles/s41598-025-97652-6)
2. Bastani et al. PNAS, 2025. Randomised trial, about 1,000 students. [Link](https://ideas.repec.org/a/nas/journl/v122y2025pe2422633122.html)
3. Fan et al. BJET. Randomised trial, 117 students. [Link](https://arxiv.org/abs/2412.09315v1)
4. Dunlosky et al. 2013. Review of study techniques. [Link](https://www.psychologicalscience.org/news/releases/which-study-strategies-make-the-grade.html)
5. Study mode pages from OpenAI, Google and Anthropic, and a press report, 2025 to 2026. Vendor and press. [OpenAI](https://openai.com/index/chatgpt-study-mode/) · [Google](https://blog.google/outreach-initiatives/education/guided-learning/) · [Engadget](https://engadget.com/ai/anthropic-brings-claudes-learning-mode-to-regular-users-and-devs-170018471.html) · [Claude Code docs](https://code.claude.com/docs/en/output-styles)