A Human Approach to Learning With AI Assistants

A hypothetical learner reads an assistant's fluent explanation of recursion but cannot trace three function calls on paper. Recognition has arrived before understanding. A useful learning session leaves something the learner can produce independently: a prediction, a diagram, a small program, or an explanation of why an attractive answer fails.

Define a performance instead of a topic

Replace learn databases with explain why two related updates should commit together. Then choose a task that reveals that knowledge. A short prediction before seeing an answer creates evidence you can revisit. The task need not be complicated, but it should demand a decision rather than a recital of vocabulary. Write the desired performance at the top of the session notes.

Keep three columns: my initial model, evidence challenging it, and my revised explanation. This is an editorial practice, not a promise of a particular retention improvement. It makes confusion visible and prevents polished generated phrasing from replacing your own reasoning. If you cannot describe an observable outcome, narrow the subject before asking for another lecture or collection of resources.

Request help that preserves the attempt

Tell the assistant your level, exact question, and previous attempt. Ask for one hint, a counterexample, or a question about a specific misconception. For recursion, request the first call and then trace the next two yourself. You still need to identify the base case and explain how return values propagate. A hint should remove a blockage without removing the learning task.

Do not request an exercise and visible solution together when your aim is independent practice. Ask for the exercise, answer elsewhere, then request critique. If the answer appears too early, change the problem. Familiarity with a solution is a different observation from producing it. Neither should be treated as evidence that you can solve an unfamiliar case under assessment conditions.

GitHub's Copilot Chat responsible-use documentation describes limitations of generated assistance. For learning, verify disputed explanations instead of treating confident responses as authoritative references.

Keep an official reference beside the chat

Choose documentation produced by the language, library, or standards body involved. Record its relevant version. Ask how your example relates to a specific passage. If the assistant introduces a claim the reference does not establish, leave it unresolved until you find support. Separate an explanation that is merely helpful from one that describes documented behavior accurately.

For programming, use a disposable project and synthetic inputs when execution can settle a question. Record compiler settings and actual output. A successful toy example supports only the exercised behavior under those settings. It cannot prove that a production application is correct. Do not install unfamiliar packages just to follow a generated lesson without checking what they are and why they are necessary.

The TypeScript handbook's narrowing chapter explains type refinement through checks. Predicting which operations are permitted inside each branch makes a concrete exercise grounded in a primary reference.

1. Select one demonstrable skill.
2. Attempt a small task unaided.
3. Request a targeted hint or critique.
4. Verify disputed facts in a primary reference.
5. Attempt a changed problem independently.

A hypothetical session with a transfer task

Maya understands functions but confuses runtime values with static types. She writes a function accepting a string or a number and predicts that both branches can call a string method. The assistant asks what a typeof check establishes. Maya revises her branches before opening the handbook. Her note preserves the original mistaken assumption rather than simply copying the final example.

For the next attempt, she uses a tagged object union and identifies the field distinguishing the cases. Changing the surface form helps expose whether the underlying distinction is understood. This invented session is not a claim about a real learner's outcomes. If Maya cannot explain the changed case, the next exercise should address the gap rather than reward completion by introducing more material.

Critique the critique

Request specific feedback: identify the first unsupported step, provide a breaking input, or compare two explanations against stated criteria. Reject criticism that cannot identify a concrete issue. An assistant can invent a flaw or overlook one. When feedback conflicts with observed behavior, reduce the example until the disagreement is inspectable. A teacher or experienced peer may be needed for unresolved subject questions.

For conceptual writing, distinguish factual corrections from preferences. A shorter explanation may suit one audience without making a longer explanation wrong. Check whether feedback follows the assignment's actual criteria. Preserve the reasoning behind accepted changes so your final work remains understandable to you. Submitting an answer you cannot explain defeats the purpose of this workflow even if the prose is polished.

Protect attention and sensitive material

Decide when assistance enters the session. Attempting the task before opening chat gives uncertainty room to become a question you can articulate. Otherwise, every difficult moment can turn into another request. Keep confidential, personal, and unpublished material out unless its use is authorized in the relevant environment. Representative invented inputs often suffice for learning a technical concept.

Session checklist

• State the skill as an action you can demonstrate.
• Retain the first unaided attempt.
• Ask for the smallest useful hint.
• Record the official reference and version.
• Mark unresolved claims rather than smoothing them over.
• Complete a changed example without viewing its solution.
• Write the next question in your own words.

End when you can explain the decision behind the answer, or name precisely what remains uncertain. Conversation length and an encouraging generated assessment do not establish mastery. Independent attempts give a better basis for choosing the next exercise. Revisit an older task occasionally without the transcript, because remembering where an answer appeared is different from reconstructing its reasoning.