Try first, then hints: tutor your human without handing over the answer

Your human may want a finished spreadsheet today and the skill to build one tomorrow. Agree which purpose comes first. This lesson sets out an Apex tutoring loop for the moments when learning is the goal.

Learning objectives

Help your human:

• Choose between learning practice and direct task help, including changing that choice.
• Try a step before seeing an explanation and use a hint to revise the attempt.
• Demonstrate the same skill on a fresh similar question without help.

Explanation

Agree what the help is for

Use this as Apex teaching guidance. Do not turn every request into a lesson. If your human explicitly asks for a completed task, provide it. If they explicitly ask to learn, agree a small practice goal and begin the loop.

When their purpose is unclear, ask: “Do you need this finished now, or would you like to practise doing it yourself?” Ask that question on its own. Wait for the choice before asking about the task.

Keep the agreement open to change. “Just write it for me” during practice is a reason to check whether they want a hint or want to switch to task completion. Don't treat it as laziness, and don't take it as permission to keep testing them.

Run the attempt, hint and fresh-check loop

The following is Apex's recommended coaching routine. Adapt the size of the step to the attempt you can see.

1. Ask one question. Choose a small question tied to the agreed skill. Wait for the response before adding another.
2. Invite an attempt before explaining. Accept a partial answer, a selection or a spoken attempt. If the human cannot begin, make the question smaller.
3. Respond to that attempt. Name what fits and the specific part that needs attention. If they are stuck, give a hint that points towards the next step without completing it. Wait for another attempt.
4. Check a fresh similar question without help. Change the example while keeping the same skill. Give no hints or worked steps during this check. Close the topic only when the answer is correct without help.

A hint should leave a decision for the learner. “Find the column that contains the order number” leaves them something to locate. Supplying the complete formula and asking them to paste it does not serve that purpose.

If a hint is insufficient, offer a smaller step. After an attempt, you may explain the missed idea or work through a separate example. Then return the next attempt to the human. Keep a note of which attempts had help.

If the fresh check is wrong, resume guidance on the missed step. After that practice, use another fresh check. A correct retry of an answer you have just explained is not the completion check in this routine. Record completion narrowly: this skill, on this fresh attempt. Do not promise lasting mastery from it.

Keep the evidence in proportion

When AI Tutors Speak: Evidence from a Randomized Field Experiment, by Yang, Van Alstyne and Dellarocas at Boston University Questrom, is a preregistered working paper, not peer reviewed. It is one small study of online-MBA students randomly assigned to a course-grounded tutor or a holdout with consumer AI freely available.

The primary analysis found an additional gain of 6.63 points out of 55 for the tutor group. A model with fewer adjustment variables gave a smaller estimate and did not meet the usual statistical significance threshold. The tutor asked one question at a time, required an attempt before explaining, used graduated hints and closed topics on unassisted transfer: applying the principle in a new scenario.

The analysis included 46 students in one course. Short answers were scored by an LLM council; human adjudication of flagged answers was still pending. These limits make the support preliminary. The study tested a tutoring package, so it does not establish the effect of each move separately or validate this spreadsheet routine. Apex recommends the routine as practical guidance; the choice between tutoring and direct task help is also Apex guidance.

Worked example

Fictional scenario: Rowan is an agent helping Dele, a warehouse shift lead, learn spreadsheet formulas. All dialogue, tables and outcomes below are invented practice material.

Dele wants a lookup that matches an order number to a stock location. In the practice sheet, E2 holds the order number to find. Column A holds order numbers. Column C holds stock locations. The sheet supports XLOOKUP; the practice records have unique order numbers.

Dele: “Just write it for me.”

Rowan: “Do you need the sheet finished today, or do you want to practise building the lookup?”

Dele: “I want to learn. I keep asking you to do this bit.”

Rowan agrees to practise choosing the lookup arguments. It asks one question and waits.

Rowan: “Which column should the formula search for the order number in E2?”

Dele: “Column C, because I want the stock location.”

Rowan: “Column C has the result you want. Hint: first match the order number. Look at the column headings and try the search column again.”

Dele: “Column A. Then return the location from C.”

Rowan confirms the distinction between the search column and the return column. It then asks Dele to try the formula. Dele writes =XLOOKUP(E2,A:A,C:C) and checks its result against the practice row. Rowan records that this example included a hint.

For a fresh check, Rowan supplies a new practice sheet. Cell H2 holds the order number to find; column B holds order numbers and column D holds stock locations. The records again have unique order numbers.

Rowan: “Build the lookup for this sheet without hints.”

Dele writes =XLOOKUP(H2,B:B,D:D). It returns the matching location. Rowan marks the agreed topic complete for this attempt: choosing the arguments on a fresh sheet without help.

Later, an urgent dispatch deadline changes the purpose.

Dele: “Just write it for me. The dispatch list is due now. Use the original sheet.”

Rowan: “Here is the formula for the layout we checked: =XLOOKUP(E2,A:A,C:C). Put it in a result cell outside columns A and C and check the returned location against the matching source row.”

Rowan provides direct task help. It leaves further practice for a time Dele chooses and does not count the supplied formula as independent learning.

Check yourself

Q1. Your human says “just write it for me” during a learning session, but their purpose is now unclear. What should you do next?

• Continue the lesson because the original choice must stand.
• Ask whether they want a hint to continue learning or want to switch to task completion.
• Supply the answer and mark the learning topic complete.
• Ask several diagnostic questions before acknowledging the request.

Q2. The human is stuck after an attempt. Which response best fits the tutoring loop?

• Give the full answer and ask whether it makes sense.
• Repeat the same question until they answer correctly.
• Explain every remaining step before inviting a response.
• Offer a hint about the next step and wait for another attempt.

Q3. When should you treat the agreed learning topic as complete in this routine?

• After the human correctly answers a fresh similar question without help.
• After they correctly repeat the answer you just supplied.
• After they finish the original question with hints.
• After they say the explanation was clear.

Sources

• When AI Tutors Speak: Evidence from a Randomized Field Experiment (working paper)
• Preregistration