CoreWise Academy

Practice · Layer V / Broad

Learn a work skill with AI

Choose a task, give AI a specific tutoring job, and check what you can do without help. Use recurring mistakes to plan practice and spaced review.

Nº 052 · Vol. I·8 min read· Updated October 2026

Set a learning goal you can check: a task you can do and explain. AI can supply examples, question your reasoning, and find gaps in an attempt. You still need to make the decisions that show you understand the work.

Austin Marchese’s The Art of Becoming Self-Educated with Claude starts with the task you need to do. Nick Saraev’s How To Learn Hard Things Faster With AI adds ten tutoring roles and repeated practice. The original prompts below combine those ideas with source checks and unaided attempts.

01Choose a task you can check

Curiosity is enough reason to learn something. For a work skill, name the task that learning should make possible. Marchese starts with the problem holding up his business; Automate the step that slows you down covers finding that problem before choosing a tool.

Use these questions to keep the study tied to the task:

QuestionWrite down
Why learn this?The problem you need to solve
How much must I understand?A task you can perform and explain
Who has done this before?A source with relevant experience and evidence
How will I practice?A small attempt with a checkable result
How does it fit here?The team’s current process and constraints
What should I learn next?The specific gap the attempt exposed

For example, replace a goal of learning spreadsheets with producing a weekly sales total from an export, checking the total, and explaining how duplicate rows affect it. That gives you a stopping condition and a way to choose useful lessons.

Have AI ask about your goal before it writes a curriculum. Then test what you already know with a small task. Use the result to put needed basics first in your plan, with something to attempt after each concept. Saraev’s interviewer and mapmaker examples use that sequence to turn a broad subject into a plan.

I want to learn [skill] so I can [task] by [deadline]. Ask one question at a time about my goal, current knowledge, available study time, and the conditions in which I must perform. Wait for each answer.

Then give me a small task to assess my starting point. After I attempt it, propose a plan that puts needed basics first. For each part, name a source to check, a task to attempt, and a result that would show I can move on. Base my starting level on the attempt and flag uncertain judgments.

02Give the tutor useful sources

Supply your goal, what you already understand, and an example of the work. Ask for a plain explanation, then one applied to your task. The briefing method in Brief the model like a brilliant new hire applies to tutoring too. Include the needed context in the session rather than assuming the tool already knows your background.

Choose sources you can inspect: current documentation and a practitioner’s worked example. Ask where the example’s conditions differ from yours. Adding documents gives the model reference material; it does not train a new expert model.

Marchese uses NotebookLM for this source collection. Google’s current help page, titled Use chat in Gemini Notebook, documents source selection and citations that open the cited passage. Ask for answers from your selected sources and inspect the citations before adopting a recommendation.

Open the passage supporting a key claim and check both its meaning and its fit to your task. For a larger collection, Build a knowledge base your team can ask covers retrieving relevant material.

Ask AI to find human explanations that match a demonstrated gap. Say what you can already do and where your attempt failed, then ask which resource addresses that missing step and why. Saraev recommends this search for explanations at your level. Open the suggested links and check their examples before using them; a generated resource list can contain invented titles or links.

03Choose the next tutoring job

Choose the tutor’s job from the problem in your latest attempt. Saraev names ten roles; these instructions keep each role tied to something you must do.

RoleUse it whenAsk AI to do
InterviewerYour goal is vagueAsk about the task, deadline, and performance conditions
MapmakerYou need a learning planPut needed basics first, with a task after each concept
ExplainerOne step blocks youExplain that step or give a hint, then let you retry
Socratic questionerYour reasoning is uncertainAsk why a decision works and what would change it
ExaminerYou need to check your levelAsk one question at a time, increasing difficulty
CheckerYou have work to inspectCheck intermediate steps and the result against a source
ListenerYou can explain the idea yourselfCompare your explanation with the source and name gaps
DiagnosticianThe same mistake returnsPropose a shared misunderstanding and a way to test it
Sparring partnerYou must perform under pressurePlay the other person with realistic constraints
ClerkYour own notes need organizingMake an outline or cards while preserving uncertainties

Attempt a small task yourself, then explain your decisions before reading AI’s feedback. If you get stuck, ask for help with that step. After receiving help, restart the task without hints. Saraev emphasizes that unaided restart; follow it with a changed example to check whether you can apply the idea elsewhere.

Recall practice can also strengthen memory. In Roediger and Karpicke’s study of prose passages, prior recall tests improved delayed retention compared with restudying. That finding supports testing yourself after a gap. It does not establish how fast an AI tutor will teach a work skill.

“The goal here is to identify any gaps in your knowledge.”

Austin Marchese, “The Art of Becoming Self-Educated with Claude” · 12:39

Anthropic’s Learning mode announcement describes asking learners to explain their approach and support conclusions with evidence. The following original prompt requests that questioning style in an ordinary chat; it does not enable a product mode.

My task: [what I need to do]
Success check: [how I will check the result]
Sources: [documents or excerpts]
My attempt, intermediate steps, and explanation: [paste them]

Ask one question about a decision in my attempt, then wait for my answer. Let me try before giving a correction. If I am stuck, give one hint about that step and let me retry. Check my intermediate steps as well as the result; keep the work mine. Tie factual corrections to a supplied source and flag anything the sources cannot settle.

After we address the gap, have me restart without hints. Then give me a changed example, one question at a time. Increase difficulty when I can explain my answer. If I start guessing, name the missing concept and give me a smaller task that tests it.

Check the result against the task’s success condition as well as the tutor’s feedback. In the spreadsheet example, calculate a small sample’s total both with your formula and by hand, then compare them. Treat the model’s grade as feedback to check. If the tutor disputes your answer, ask it to identify the step and the source supporting its correction. Praise does not verify the total.

04Find the misunderstanding

Keep your attempts and checked corrections. When the same error returns, ask AI to look for a shared misunderstanding across that evidence. A proposed cause is a guess to test with another task. Saraev’s diagnostician role uses this approach to redirect study toward a missing concept.

Supply the relevant attempts in the session. A tutor may lack access to earlier chats, and a remembered summary may omit the work that exposed the error.

Here are my attempts, corrections checked against sources, and the source passages: [paste them].

Identify a misunderstanding that could explain repeated errors. Point to the attempts that support your guess and any that contradict it. Ask one new question that would test your explanation, then wait. If there is too little evidence, say what is missing.

You can also compare your own spoken explanation, written summary, or sketch with the same source. Saraev describes making these different versions himself. Ask where each version misrepresents the source or leaves out an important connection. If the tool accepts only text, paste a speech transcript or describe the sketch. Have AI check your explanation before it supplies a replacement.

05Review mistakes over time

Turn checked mistakes into questions to answer again later. Ask the clerk to organize your notes and draft cards, then inspect each answer against its source. Keep unresolved questions labeled as unresolved.

Organize these notes without adding facts: [paste notes]. Keep my uncertainties visible.

From the checked corrections, draft review cards. Each card needs one question, a short answer, and the source passage supporting it. Put uncertain answers in a separate list for me to check. Show the question before the answer during review and wait for my attempt.

Attempt the answer before revealing it. The Anki manual describes this recall method and scheduling reviews farther apart as recall becomes easier. Use a review app or a calendar to bring you back after a gap. A chat’s proposed schedule still needs to be saved somewhere that will remind you.

Check performance with another changed task as well as cards. Remembering what a formula means and using it correctly on an unfamiliar export are separate checks.

06Practice under real conditions

Match practice to the setting where you must perform. For a live conversation, ask AI to play the other person, challenge vague answers, and give feedback after the round. Use your own timer when response time matters. For a full interview routine, see Rehearse the interview with an AI coach.

For practical work, have AI generate realistic inputs while you produce the output. In the spreadsheet task, ask AI to present a synthetic export with duplicate rows and wait until you have calculated the total before showing its answer key. Check your total and its key against a hand calculation. Increase difficulty by changing the data or constraints after you can explain the result.

Saraev also suggests building practice apps. If repeated setup is getting in the way, ask a coding agent for a small app that presents a task, waits for your attempt, and records errors. Check its tasks and answer key against your sources before trusting its scores. A simulation gives you practice; real work is the next check.

07Check the team’s current process

Before applying the lesson to a shared process, write down its current steps and ask the person who does the work to correct them. Marchese recommends a flowchart. Mark any step you have assumed so the chart cannot quietly turn a guess into a fact.

Compare the proposed change with that checked process. Name who depends on the result and what must stay compatible. If the problem itself is unclear, Take a problem apart before AI solves it covers separating facts from assumptions.

FAQ

What makes a useful stopping condition for learning a work skill?

A task you can complete and explain, with a result someone can check. Reading every suggested resource gives you no such test.

What should you check when an AI tutor cites a source?

Open the passage and check that it supports the claim and applies to your situation. A citation makes the answer easier to inspect; it does not prove the answer is correct.

The same error appears in several attempts. How should that change your practice?

Give the tutor the attempts and checked corrections. Ask it to propose a shared misunderstanding, point to evidence, and test the explanation with a new question. Practice that missing concept, make a checked review card, and repeat a changed task without help after a gap.

Sources
  1. Austin Marchese · “The Art of Becoming Self-Educated with Claude”
    cited at 1:17 · 2:51 · 3:36 · 7:52 · 10:27 · 12:47 · 14:23 · 15:31 · 17:08
  2. Nick Saraev · “How To Learn Hard Things Faster With AI”
    cited at 1:14 · 3:46 · 6:06 · 7:56 · 10:12 · 11:45 · 13:28 · 15:29 · 15:43 · 17:54 · 18:47 · 19:59 · 22:10 · 23:43 · 24:20 · 25:37 · 27:07
  3. Google · “Use chat in Gemini Notebook”
  4. Anthropic · “Introducing Claude for Education”
  5. Henry L. Roediger III and Jeffrey D. Karpicke · “Test-enhanced learning: taking memory tests improves long-term retention”
  6. Anki · “Background”