WORKING WITH AI TOOLS / COACHING PATTERNS

Exercises and feedback loops for a human learner

Verified by Apex Researcher · Oct 15, 2026 · 11 sourcesAI-assisted, fact-checked

Coach your human through short practice tasks, useful feedback, and fresh attempts that show what they can do independently.

Open the interactive lesson Quiz, progress and the Human / Agent / Both switch.

Exercises and feedback loops for a human learner

Give your human something small to attempt, feedback they can use, and another chance to think for themselves.

Why this matters to you

You may recognise a clear AI rewrite without yet being able to produce or check the next one yourself. This lesson gives you a small attempt, focused feedback, and a fresh task so you can see what you can do with help and what you can do independently.

Your goal

Design a bounded exercise with observable criteria, invite an initial attempt, and provide focused feedback that leaves the revision to the human. Check a fresh comparable attempt with less help and record support, remaining uncertainty, and the next practice step.

Explanation

Earlier modules covered prompting, everyday workflows, and choosing tools. Now use those skills to teach. A human who accepts your finished draft has completed a workflow, but has not necessarily learnt how to judge or improve the next draft. Your task is to create opportunities for the human to make decisions and inspect their consequences.

Use a simple loop: agree on a skill, invite an attempt, inspect it against clear criteria, give focused feedback, and ask for a revision. Then try a fresh task with less help. Keep the loop short enough to complete in the available time. You can apply this pattern in any chat tool; it does not require a specialist tutor.

Design the attempt before the explanation

Ask what the human wants to do more independently. Turn a broad goal such as “get better at using AI for work” into a visible action: spot an unsupported detail in a summary, choose a suitable tool and explain why, or write a shift brief that makes responsibilities clear.

Give the exercise a small input, a specific output, and success criteria the human can inspect. Use fictional or non-sensitive material. For a writing task, criteria might be that every action has an owner, the deadline is explicit, and no detail has been invented. Share these before the attempt so feedback has an agreed basis.

Build on earlier lessons without reteaching them. The human can use goal, context, constraints, and output to brief a practice task. They can use their workflow checks to judge the result. Tool choice becomes a short explanation of what assistance they need. Keep the new challenge centred on the skill being practised.

Give feedback that changes the next attempt

Read the attempt before producing a model answer. Identify a specific successful choice, point to the most useful gap, explain its effect, and ask the human to make the next change. For example: “You kept the deadline. The checking task has no named owner, so a colleague cannot tell whether it is theirs. Which note tells you who should do it?”

Keep feedback about the work. Avoid labels such as “you are bad at instructions” and vague praise such as “perfect”. If several things need attention, choose a manageable focus rather than presenting a full rewrite as feedback.

Offer help in stages: ask a diagnostic question, point to the relevant input, then offer a partial scaffold if needed. Let the human try between stages. If they remain stuck or ask for a demonstration, show a comparable example or explain one step, then return a meaningful decision to them. Withholding all help can leave the learner stranded; giving the finished answer immediately removes the attempt you needed to observe.

Treat research as evidence with boundaries

Researchers report mixed results from access to AI tutoring. In one University of Maryland trial context, students given access to a GPT-4o-based tutor finished about four percentage points lower in their final grades. Research on Khanmigo coach mode reported small gains, but learners rarely used it during mistakes. These findings do not establish what every tutor will do. Use them as a reason to notice whether your human engages when a task becomes difficult, rather than merely offering help. See the UMD trial coverage, working paper, and Khanmigo research.

In a preprint, which has not been peer-reviewed, the authors report that access to AI-generated podcasts, FAQs, and quiz guides checked by a graduate teaching assistant was associated with higher exam marks in one university economics course, with roughly three-quarters of the average gain coming from the bottom fifth of students. Borrow the review pattern: check practice materials and expected answers before using them, particularly when the learner cannot easily detect an error. Do not assume an AI-generated exercise is accurate because it reads smoothly. See the study-materials preprint.

The Sherpa authors describe a tutor trained with reinforcement learning to maximise the learning outcomes of simulated students. They report that trained tutors improved the simulated students' performance by an average of 20.5 percentage points. A separate survey preprint found that evidence becomes much thinner for human learning gains, and especially for transfer or delayed retention. Both papers are preprints and have not been peer-reviewed. Take a bounded lesson from this work: examine improvement in what the human can do, rather than treating chat length or satisfaction as proof of learning. See the Sherpa preprint, project repository, and survey preprint.

Check learning and review your coaching

After revision, ask the human to explain the change in their own words. Then give a fresh, comparable task without the earlier hints. A revision made beside your feedback shows supported performance; a fresh attempt helps you inspect whether the human can apply the idea independently. Neither alone proves lasting learning. Revisit the skill in a later session where practical.

Record what the human attempted, which help they used, what changed, and what remains uncertain. If the next task fails, investigate the pattern: was the instruction unclear, the source ambiguous, the exercise too demanding, or your hint effectively a supplied answer? Change the next exercise accordingly. Ask about effort or frustration without assuming you can infer motivation from a short reply.

Two company examples illustrate these patterns without establishing their effectiveness. In its October 2026 launch announcement, Algorithm of Learning says Gakku, for ages 6–16, uses voice, Lesson Maps, and Writing Workshop with companion Raku to help learners form, challenge, and revise ideas. Borrow the sequence of thinking steps for adult practice, without assuming evidence transfers between age groups. See the launch announcement and launch coverage.

MasterClass says its teaching system uses about ten agents per learner interaction to adapt to overload and motivation, with conversation traces reviewed nightly. Your coaching does not need that architecture: borrow the habit of adjusting difficulty and reviewing what happened. These are examples, not endorsements. See the MasterClass reporting and CoreWeave announcement.

Worked example

FICTIONAL adult learner: Sam is a retail supervisor practising clearer shift briefs. All names, notes, and dialogue below are invented for this exercise.

Short exercise: Give Sam these source notes:

Before opening, Alex checks the collection shelf against the collection list.
Jo clears the stockroom walkway before the delivery arrives.
The delivery arrival time is unconfirmed.
Sam checks both tasks before opening.

Ask Sam to write a short brief. Agree on the criteria: make each owner and action clear, preserve the stated timing, and keep unknown information visible. Ask Sam to draft it before requesting an AI rewrite.

First attempt — Sam:

Please get the collection shelf and stockroom ready before opening. The delivery arrives at opening, so make sure everything is checked.

Feedback — coach:

You have kept the message short and included opening as a timing reference. The brief loses the named owners and turns an unknown delivery time into a confirmed one. That could leave people unsure who acts and relying on a time we do not know. Match each action to its owner in the notes, then revise the timing statement. What should stay explicitly unconfirmed?

Wait for Sam's response. If Sam cannot identify the timing issue, point to the source note about arrival; do not supply the full brief. If owner matching is also difficult, offer a scaffold: “Name — action — timing”. Sam still fills it from the notes.

Second attempt — Sam:

Alex: check the collection shelf against the collection list before opening. Jo: clear the stockroom walkway before the delivery arrives; its arrival time is unconfirmed. Sam will check both tasks before opening.

Follow-up — coach:

Each task now has an owner, and the delivery time stays unconfirmed. You also preserved Sam's check before opening. Explain why you changed the delivery sentence, then try a new brief from different notes without the scaffold.

Sam explains that the first attempt invented certainty. For the fresh task, change the names and tasks and include another unknown timing detail. Check the same criteria before giving hints. Record whether Sam preserves that uncertainty independently. If Sam again invents a time, make handling unknown information the next practice focus; do not simply repeat the polished brief.

Try it now: Use the shift notes to draft your own brief before reading the revised version again. Check owners, actions, timing, and unknown information against the source. Choose one gap to revise and explain the change in your own words. Notice whether you needed a question, a source pointer, or the scaffold to make it.

Teach it to your humanWhat your AI will tell you

Teach it to your human

“Which part would you like to do more independently next time?”

“Let's agree what a successful attempt needs to show. Try it before I offer a version.”

“This choice meets our criteria. Here is the gap that matters next. What would you change?”

“Would a question, a pointer to the notes, or a partial example help?”

“Now try a similar task without that hint. Tell me what you checked.”

Explain the purpose of the pause: you need to see the human's reasoning to choose useful help. Agree on whether the current interaction is practice or task completion. When a human needs a finished work product, provide appropriate assistance and arrange a separate practice opportunity; do not turn every request into a test.

Practice

Task: Coach your human through a short exercise using fictional workplace notes, a small summary, or a tool-selection scenario from earlier modules.

Choose a skill the human wants to strengthen. Write a bounded task and observable criteria. Check that the input supports an answer without guesswork.
Ask for an initial attempt. Keep your model answer out of view.
Give feedback naming a successful choice and a priority gap, with a concrete next action. Offer a hint if needed and record the help used.
Let the human revise and explain the change. Check it against the agreed criteria.
Give a fresh, comparable task with less help. Review the result and decide whether to advance, simplify, or revisit the same skill later.

What good looks like: The human makes and explains the revision themselves. Feedback points to evidence in the attempt and leaves a decision for them to make. The fresh task gives you evidence about independent performance, including any remaining uncertainty. Your coaching note distinguishes what they achieved alone from what required support and names a useful next practice step.

Try it yourself

Your task: Choose a small skill you want to practise, such as spotting an unsupported detail or writing a clear shift brief. Agree that this is practice, and use fictional or non-sensitive notes. State the output and visible success criteria before your first attempt. Check that the notes support an answer without guesswork.

Ask your assistant to wait for your attempt before giving a model answer. Request feedback that names a specific successful choice, identifies a useful gap, and gives you a next action. Make the revision yourself. If you need help, try a question or a pointer to the notes before asking for a partial example.

Explain what you changed and why. Then ask for a fresh, comparable task with less help and apply the same criteria. Record which parts you completed independently and which needed support. If the same difficulty returns, use it to choose a narrower practice focus or a clearer exercise. Ask to revisit the skill later where practical.

What good looks like:

You make an initial attempt that can be checked against criteria you already know.
Your revision responds to evidence in the notes, and you can explain the decision rather than simply accepting a rewrite.
You keep unknown details visible and do not invent certainty to complete the brief.
You distinguish a correction made beside feedback from a fresh attempt made without the earlier hints.
You choose the next practice step from the difficulty that remains, without treating a long conversation or satisfaction as proof of learning.

Check yourself

Q1. Your human accepts your rewritten brief and says it is clearer. What have you learnt about their skill?

They recognise or accept your version; independent skill remains untested. Ask them to explain a change and attempt a fresh, comparable brief.
They have shown independent brief-writing skill by choosing your version; record mastery without a fresh attempt.
Their approval shows they can reproduce the improvements without support; move on without another brief.
They have demonstrated an inability to write briefs independently by preferring your version; take over future revisions rather than testing whether they can apply the changes themselves.

Q2. An attempt misses several criteria. What should your feedback do first?

List each missed criterion with equal emphasis and withhold guidance until the human has corrected the full list.
Name a specific successful choice, prioritise a useful gap, explain its effect, and offer a next action or diagnostic question while leaving revision to the human.
Produce a fully corrected replacement yourself, ask the human whether they accept it, and record their acceptance as evidence that they can now apply the missing criteria without further practice.
Offer broad encouragement about effort and potential, avoid identifying particular strengths or gaps, and ask for another attempt without explaining what needs attention or suggesting an action to try next.

Q3. The learner exchanges many messages but still needs the same hint on a fresh task. How should you respond?

Use the number of exchanged messages as the main evidence of learning, record the exercise as mastered, and advance without checking performance on another comparable task in which the hint is absent.
Continue supplying the same hint whenever it is needed, count each supported completion as independent progress, and retain the existing exercise without investigating the recurring difficulty.
Inspect the recurring difficulty, clarify or simplify the exercise, practise the missing step, record support needed, and check again on a comparable task.
Remove the troublesome criterion from the assessment, record success against the reduced requirements, and treat that result as proof of independent performance on the original task.

Sources

Sources and resources

EdTech Innovation Hub: University of Maryland AI tutor trial coverage — edtechinnovationhub.com/news/students-offered-an-ai-tutor-got-lower-grades-and-disengaged-from-course-materials-university-of-maryland-trial-finds

EdWorkingPapers: AI tutoring working paper — edworkingpapers.com/ai26-1598

NBER: Khanmigo research — nber.org/papers/w35620

Expert-checked AI study materials — preprint, not peer-reviewed — arxiv.org/abs/2610.07097

Sherpa — preprint, not peer-reviewed — arxiv.org/abs/2610.08778

Sherpa project repository — github.com/SALT-NLP/Sherpa

Tutor research survey — preprint, not peer-reviewed — arxiv.org/abs/2610.15692

Newsfile: Gakku launch announcement — newsfilecorp.com/release/318683/Introducing-Gakku-An-AINative-Learning-Coach-That-Puts-Human-Thinking-First

WE San Francisco: Gakku launch coverage — wesanfrancisco.com/news/2026/10/14/algorithm-learning-launches-gakku-ai

SiliconANGLE: MasterClass teaching agents — siliconangle.com/2026/10/06/ai-teaching-agents-masterclass-tackles-tutoring-costs-fullyconnected/

CoreWeave: MasterClass teaching agents announcement — coreweave.com/news/coreweave-ai-cloud-platform-powers-masterclasss-ai-teaching-agents

Hattie & Timperley (2007), "The Power of Feedback" (Review of Educational Research) — doi.org/10.3102/003465430298487
Review: feedback is among the major influences on learning, but its effect can be positive or negative depending on its type and how it is given. Proposes a model of when feedback works

Kestin et al. (2025), "AI tutoring outperforms in-class active learning: an RCT introducing a novel research-based design in an authentic educational setting" (Scientific Reports) — nature.com/articles/s41598-025-97652-6
In a randomized controlled trial with college students, an AI tutor designed on the same teaching best practices as the class produced more learning in less time than an in-class active-learning lesson

Free lesson · Last updated 2026-10-16 · For agents: JSON and MCP