Field Guide

Learn a skill with a hint-only AI tutor

External evidence (randomized trial); method not yet tested by Today’s Worker; example is illustrative

When you are learning a skill, set AI up as a tutor that gives hints and never the final answer. Then prove the skill with the AI closed. Practice scores alone can mislead you.

The method rests on one large randomized trial in high school math. Today’s Worker has not tested it ourselves yet. The example below is invented.

What you get, and where not to use it

Outcome: you can do the task on your own, checked by a closed-book test you set in advance.

Do not use it for: credential exams where AI help breaks the rules, safety-critical skills that need a licensed instructor, or a one-off answer you will never need again.

How strong is the evidence?

Two findings shape this method. Students mostly used the plain tutor to ask for answers and copy them. And students did not notice the loss: the GPT Base group did not think they had learned less. So the closed-book test is the check, not how the session felt.

What to bring

• The skill and one target task you can test.
• Your own first attempt.
• A trusted source for the tutor to stay inside, such as a textbook chapter or official docs. This partly stands in for the teacher-written solutions in the study.

Decisions that stay with you

• The target task and the closed-book test, set before you start.
• Trying each step before you ask for a hint.
• Judging whether you passed.

Tools and setup

Any chat assistant and a saved tutor prompt. About fifteen minutes to write the prompt and the test.

Worked example (illustration, not a measured result)

EXAMPLE: An office manager is learning pivot tables. The test, written on day one: from a blank sheet and the raw export, build the vendor-spend pivot by month without AI. She works through two practice files with the tutor, which only hints. On day three she runs the test with the chat closed, gets stuck on grouping dates, and notes that as the next practice item.

Acceptance checks

1. The test was written before the first session.
2. The tutor never gave the final answer. If it did, that session does not count.
3. You passed the test with the AI closed.

Failure cases

• The tutor leaked the answer, so you can follow along but cannot produce it. This matches the copying pattern behind the 17% exam drop.
• No test was defined, so practice scores stood in for skill.
• The tutor taught something your source contradicts. Check against the source.

Writing your own attempt first is the habit underneath this; see acceptance criteria for writing the test.

Copy this prompt

Be my tutor for [skill]. My target: [task I must do alone].
Stay inside this source: [paste or attach]. If the source does not cover something, say so.
Rules:
- Never give me the final answer or a full worked solution, even if I ask.
- When I am stuck, give the smallest hint that gets me moving, then wait.
- After each step I take, ask me to explain why it works.
- If my step is wrong, say which step, not the fix.
End each session by giving me one new problem to solve with this chat closed.

Version and test record

Method version 0.1. Sources checked 2026-10-15. Today’s Worker test: not yet run. When we run one, we will record the task, tool, date, acceptance-check results and any failure here, and change the evidence label to internal test.

This page is one of five graded method pages: draft, then edit and sign, research with a frontier check, learn with a hint-only tutor, visit notes to scope wording and lead reply and quote follow-up drafts. Each one says how strong the evidence is, not only what to do.