Chapter 4: The Question Is the Unit of Learning

Turn a selected uncertainty into a portable inquiry brief, state your own expectation before receiving assistance, and judge the answer through an explanation and an observable attempt.

A better-sounding question can still miss the problem

Audy opens the curiosity map she made in Chapter 3. Her selected inquiry concerns a proposed workshop in which adult learners turn meeting notes into an action list. The example is fictional, including its audience and thirty-minute duration. It gives her a small task to design before she makes claims about Himura Inc.'s usefulness to real educators.

Her earlier inspection showed what she herself would check. Now she wants workshop-brief-v1 to guide a fresh draft without relying on her memory of that inspection. Her first wording is, “How can I create an engaging, effective, learner-centered workshop that helps people take action after meetings?” It sounds more considered than her original request for a template. It contains several worthwhile aspirations. It also invites an answer so broad that she will struggle to tell which part solved her problem.

The assistant returns a lively agenda: an opening story, a discussion, a worksheet, a paired exercise, and a reflection. Audy could use the response to fill thirty minutes. Yet a learner could participate enthusiastically without ever distinguishing a firm commitment from a suggestion. The answer has improved the presentation of an unspecified result.

She rereads the map. The narrower question was whether the proposed activity would produce an action list that someone could check against the commitments in the notes. The provisional criteria already exist; the activity must give each learner a chance to apply them. That question survives outside the chat because it points to an input, an output, and a judgment. It asks her to resolve something specific before she expands the workshop.

Audy replaces the request with an inquiry brief. She names the task, supplies synthetic notes, writes a provisional answer, and defines what she will inspect. The resulting document can guide an assistant, a conversation with a peer, or her own work next week. Its value does not depend on finding a magical phrase.

Bring your selected question from the curiosity map. A portable brief will test whether its apparent precision reaches the evidence and decision, or whether it still needs narrowing.

Make the question small enough to carry work

People learn through observation, imitation, repeated action, and many other experiences. This book uses a question as a manageable unit for connecting an uncertainty with an attempt to resolve it.

A useful question has a before and an after. Before investigating, you can describe what you do not know and what you expect. Afterward, you can describe what changed, what evidence produced the change, and what remains uncertain. If neither side is visible, an impressive answer may leave you without a clear account of what you learned.

Consider “Tell me about workshop assessment.” The response may contain useful concepts, but the request does not identify the decision you need to make. “How can I tell whether this action-list activity gives learners practice in distinguishing commitments from suggestions?” directs attention to a particular design. The answer could change the activity, its instructions, or the proposed check.

Small does not mean trivial. A narrow question can expose a major problem. If Audy discovers that the activity practices formatting rather than identifying commitments, she may need to revise the entire brief. The inquiry stays small because she can inspect the mismatch without redesigning every part of Himura Inc.

You can preserve a large ambition while breaking the investigation into questions that produce evidence. “Can I build a sustainable business?” remains important. It becomes workable through inquiries about a particular customer’s task, a proposed result, a delivery method, and the resources required. Each answer earns its place by affecting a decision.

Describe the situation before asking for the answer

The same words can describe different problems. An action list for a project team may need formal approval. A personal reminder list may be useful without it. A workshop exercise may deliberately include missing information so learners can practice noticing the gap. A response that does not know which situation applies may confidently solve the wrong one.

Audy’s brief identifies the immediate purpose: improve a synthetic workshop-design example. It names the learner task: convert short meeting notes into a list of agreed actions while preserving uncertainty. It names the output she needs now: proposed practice instructions and a check that can distinguish an acceptable action list from a plausible but inaccurate one.

She gives the inquiry a clear scope: inspect the design. Learning gains, demand, and the suitability of a thirty-minute duration will need different evidence.

Then she describes the available material. She has her own provisional brief, a set of invented notes, and the claims ledger. She has no recorded learner attempts or educator interviews. The assistant therefore has permission to help design an exercise, but no basis for reporting how customers or learners actually behave.

Write this situation in ordinary language. You do not need specialized formatting. A short paragraph that names the task and available evidence is often more useful than a long role description. “You are a world-class expert” cannot supply facts about a workshop that has never occurred.

If your material includes someone else’s information, use it only in ways you are authorized to use it. The essential exercise works with invented examples or your own non-sensitive work. You can learn to define the inquiry before connecting it to real records.

Write what you think before assistance arrives

Audy already has provisional criteria from her own inspection. Her first hypothesis for the next attempt is that clearer outcome wording will be enough to guide a fresh draft without carrying those criteria into the instructions. She writes: “If I replace ‘understand effective follow-up’ with ‘produce an action list that preserves agreed commitments and flags missing details,’ the new draft will retain the distinctions I identified.” This is a proposed relationship, not a result.

Writing the expectation makes it possible to discover that she was partly wrong. Without it, she could read the assistant’s response, accept a different explanation, and later remember that explanation as what she had intended all along. The record gives her a visible starting point.

A hypothesis need not sound scientific. It can be as simple as, “I think the missing owner causes this confusion,” or “I expect this input to produce a request for clarification.” State what you expect to observe if your explanation is useful. Also identify a result that would make you revise it.

For Audy, a challenge would be finding that the clearer outcome still permits an activity with no actual action-list attempt. If the agenda only asks learners to discuss good practice, improved wording has not yet changed what they do. Another challenge would be a check that rewards tidy formatting while missing invented commitments.

Give a competing explanation a fair chance. Perhaps the problem lies in the practice input, which contains no ambiguous examples. Perhaps the criteria are too vague. Perhaps Audy herself has not distinguished a suggestion from an agreement consistently. Each possibility leads to a different repair.

Keep the first expectation even when it turns out to be incomplete. Comparing it with the result makes the change in your understanding visible.

Ask for the evidence that could change your mind

Before gathering information, decide what kind would matter. Audy can inspect whether instructions require a learner to produce an action list. She can compare a proposed answer against the supplied notes. She can introduce an error and see whether her criteria identify it. These activities can expose defects in the design.

They cannot tell her whether learners will improve after the workshop. That question requires evidence about learners’ performance under relevant conditions. A source about a learning technique might help her design such an investigation, but it cannot stand in for an outcome from her own workshop.

Her inquiry therefore has a deliberately modest decision: should the brief retain its revised outcome alone, or also require a corresponding practice input and checking criteria? She can make progress on that decision with the materials she has. She leaves the learning-effect question in the claims ledger as unresolved.

Be specific about what would count against your preferred option. If every possible response can be interpreted as support, the inquiry is not giving you a meaningful test. “The assistant liked my idea” is especially weak when you supplied the idea and asked for reasons to use it. Ask for a concrete counterexample or failure case instead.

Audy chooses a challenge case in which the notes mention a possible action without recording agreement. A generated list that treats the possibility as a commitment should fail her check. This is a testable property of a document, and she can explain why it matters without appealing to the assistant’s authority.

The check belongs in the inquiry brief before the final answer is produced. Otherwise, you risk accepting whatever the response happens to do well. An output can be eloquent, comprehensive, and still miss the requirement that motivated the work.

Work through a miniature example

Audy creates three lines of fictional meeting notes. The facilitator agreed to circulate the notes by Friday. The group suggested asking the venue about access arrangements, but assigned nobody to do it. The group agreed to try a shorter introduction at the next session, without assigning an owner. These lines contain an explicit commitment, a suggestion, and an agreement with missing information.

Her first practice instruction is, “Turn these notes into a clear action list.” She can already imagine a polished but faulty answer: “Facilitator: circulate notes and confirm access by Friday. Organizer: prepare a shorter introduction before the next session.” The response supplies an owner for the third item and quietly attaches the venue inquiry to the facilitator’s Friday deadline.

A reviewer might miss the changes if the list looks reasonable. The invented details are plausible in ordinary work. That is why Audy’s check cannot depend on whether an entry sounds sensible. It must compare each assertion with the notes.

She marks the first item’s circulation commitment as supported. The added access task is unsupported as a commitment, and its owner and deadline were not supplied. The shorter introduction was agreed, but its owner remains unknown. A better list would keep the agreement while exposing the missing assignment. The venue suggestion belongs under an unresolved question unless later evidence establishes agreement.

This inspection revises her hypothesis. Clearer outcome wording helps, but does not by itself specify what an acceptable attempt looks like. The brief also needs examples containing meaningful ambiguity and criteria that distinguish supported details from invented ones.

Audy then changes the practice instruction: identify the agreed actions; preserve an owner or deadline only when stated; separate suggestions from commitments; and flag information needed to make an agreed action usable. She keeps the notes unchanged. Keeping the input stable helps her see what changed in the instruction rather than attributing every difference to a new example.

She makes a fresh action list and reviews it against the notes. It now preserves the circulation commitment, leaves the venue question unresolved, and marks the introduction’s owner as missing. That is a better result for this example. She records the improvement without calling it a controlled experiment or proof that the wording will work with every set of notes.

One more check matters. Does the practice actually ask learners to make these distinctions themselves, or does it show them a completed list and ask whether it looks clear? The latter may be a useful discussion, but it does not demonstrate the same performance. Audy adds an individual attempt before the comparison. Her brief now connects the intended task, the learner’s action, and the evidence she plans to inspect.

Treat a prompt as a reusable instruction

The inquiry brief is more durable than the conversation that helps answer it. It can contain the question, context, provisional explanation, allowed evidence, requested output, and acceptance criteria. Those fields remain useful when you change assistants or ask a colleague to review the work.

Audy saves it as inquiry-brief-v1. She includes the synthetic notes directly or links to a file she controls. She avoids references such as “use the thing we discussed earlier” because they will become unclear when the document leaves the chat. If a source is necessary, she names it and explains which claim it supports.

Her request is bounded: inspect the draft against the notes, identify unsupported additions, and propose the smallest revision that makes the practice checkable. Each proposed change needs a concise explanation tied to the source.

The portable inquiry prompt provides a reusable version. Copy it into a chat if you want assistance, or use its fields as a checklist for manual review. The content should make sense to a person who has not seen the preceding conversation.

If the assistant requires information you do not have, its useful response may be a question or a qualified suggestion. Specify that behavior in the brief. “Leave the owner unknown when the notes do not name one” is more actionable than “be accurate.” Examples of acceptable uncertainty help prevent a demand for completeness from encouraging fabrication.

A portable instruction also makes disagreement easier to examine. If two reviewers reach different conclusions, ask which input, criterion, or assumption produced the difference. You now have a shared object to revise instead of a dispute about which response sounded more convincing.

Separate receiving an answer from learning it

After the comparison, Audy can repeat the phrase “preserve uncertainty.” That is useful vocabulary, but it does not yet show that she can apply it. She needs to recognize a new instance in which a detail should remain unresolved.

Research on retrieval practice supports making room for an attempt to produce what you know. In two experiments with prose passages, Roediger and Karpicke found that prior testing improved delayed retention relative to repeated study, although repeated study performed better on the five-minute test. These were student memory experiments, not tests of workshop entrepreneurship or this chapter’s procedure. (Roediger & Karpicke, 2006)

The book adapts the principle through a practical sequence: attempt an explanation, check it, and return to it later. Close the assistant and explain the distinction you just used; then compare your account with the source and example to find what needs further practice.

Audy explains that a suggestion describes a possibility, while a commitment records an agreement to act. An action may be agreed without having every execution detail assigned. She should therefore avoid solving missing information by inventing it. She compares this explanation with her notes and corrects any overstatement.

Later, she can repeat the check using a different example. The interval is a practical choice, not a validated schedule for this book. What matters is that she does more than reread a familiar answer and mistake its familiarity for an ability she has demonstrated.

Make the unaided check useful rather than punitive

A thinking-without-AI check should expose the judgment you need for the task. It should not become a demand to memorize a whole reference library. Audy needs to distinguish an unsupported assertion from a supported one. She does not need to recite the names of every possible assessment method before improving her brief.

She creates a new synthetic sentence: “We could send a reminder on Thursday; a decision will be made tomorrow.” Without consulting the earlier output, she describes how it should appear in the action record. If she immediately assigns Thursday as a confirmed deadline, she has found a misunderstanding worth repairing.

She then introduces a plausible error into her own list and tries to explain the defect. This is a different challenge from producing a correct list from scratch. In later chapters, reviewing an agent’s output will require noticing precisely this kind of reasonable-looking mistake.

There is also reason to distinguish assisted success from independent performance when designing learning with AI. In a field experiment with high-school mathematics students, Bastani and colleagues found that an ordinary GPT-style interface could improve assisted practice while harming subsequent unaided performance; a safeguarded tutor largely mitigated that harm. The population, subject, and interventions differ from our exercise. The finding supports taking tool design and independent checks seriously, not declaring all assistance harmful. (Bastani et al., 2025)

If you fail your check, reduce the task. Compare one supported sentence with one unsupported sentence. Explain the difference, consult the evidence, and try another pair. Keep the error as a guide to your next practice.

Diagnose the question when the answer disappoints

An unsatisfactory answer does not always require a more elaborate prompt. Sometimes the question asks for a conclusion the evidence cannot support. Sometimes it combines several decisions. Sometimes the requested output is incompatible with the conditions you supplied.

Suppose Audy asks, “Will this workshop help independent educators build confidence and improve their learners’ workplace performance?” Her synthetic notes cannot answer that. More detail about her fictional workshop would not turn the question into a measured outcome. She should preserve the question and design a future evidence-gathering activity, while asking a smaller design question now.

Suppose instead that the assistant keeps giving agendas when she needs checking criteria. The question may be clear, but the requested output is not. She can repair the output instruction: return a small set of criteria and illustrate each using the supplied notes. She does not need to rewrite the whole project description.

Another failure occurs when the question contains its own answer. “Why is a task-focused workshop better than a generic template?” encourages a defense of the preference. Audy can replace it with, “Under what conditions would each approach be useful, and what evidence is missing for choosing between them?” A generic template may still help an educator whose main problem is organizing already well-defined material.

The repair should match the defect. Narrow an overloaded question. Supply missing context. Change an output specification. Find a different evidence source. Mark an unanswerable claim unresolved. A long prompt assembled from unrelated warnings can make the real requirement harder to find.

Keep the result connected to its limits

Audy ends the inquiry with a concrete revision. workshop-brief-v1 now asks for a stated learner task, an attempt that exercises it, and criteria connected to the task’s important distinctions. The action-list example demonstrates why those elements need to fit together.

In claim-evidence-ledger-v1, she records what the exercise supports: the revised instructions and checks help her detect specified defects in the synthetic example. She also records what remains open: whether the design is appropriate for real learners, whether an educator needs this assistance, and whether Himura Inc. could deliver it economically.

She keeps the rejected output because it explains the new requirement. “Do not invent deadlines” has more meaning when paired with the example in which an unrelated deadline spread to another action. Future versions of the prompt can be tested against that failure without reconstructing the whole conversation.

Her inquiry brief also contains a next question: would a different set of ambiguous notes expose a distinction the current criteria miss? This is a more useful continuation than “learn more about assessment.” It grows from the result and asks for evidence that could challenge it.

The original question, attempt, correction, and remaining boundary travel together. The same record can explain a decision to keep a process manual, abandon a design, or seek informed help.

Adapt the inquiry to the result you owe

A service professional might investigate whether a proposed project summary distinguishes requests from commitments. The input is an authorized or synthetic conversation record. The output is a summary with traceable decisions. A useful challenge case includes a suggestion that nobody accepted. Your acceptance criteria should expose the difference before the summary is sent to someone else.

An educator might investigate whether a practice activity prepares learners for the kind of decision they will face later. Ask learners to attempt the decision when appropriate and permitted, rather than relying only on their reactions to the explanation. Keep the design judgment separate from any claim about measured learning. AI can help create contrasting examples; a qualified human still needs to check them.

A software builder might ask whether a user can tell that a record was saved. The inquiry can begin with a paper walkthrough showing successful, failed, and interrupted attempts. A message saying “saved” is not enough if the record does not exist afterward. You can define that requirement before choosing a programming language or asking an agent to implement it.

Each adaptation points toward an inspectable result. Choose the artifact and reviewer your domain requires, then check whether you can explain the judgment on which the result depends.

Field assignment: create a brief that travels

Complete the inquiry brief with your selected question. Include the situation, decision, starting expectation, evidence route, requested output, and a result that would change your mind. Keep the first version before revising it. The difference between versions may be the clearest record of what the exercise taught you.

Attempt the inquiry using a no-code method: a document, a small example, and a direct comparison. An assistant is optional for the first attempt. If you use one, provide the same brief you could give a person, and keep the response alongside your own review. Do not replace the review with a favorable rating from the same model that generated the work.

Your artifact passes this chapter’s practice check when another reader can explain the question without seeing the chat, inspect the evidence you used, identify a supported revision, and find the unresolved claims. You should also be able to explain one important distinction without consulting the assistant and recognize a deliberately introduced error in a fresh example.

If you cannot do those things, choose the smallest repair. Add the missing input. Separate two questions. Replace a vague criterion with a contrast. Revisit the source. Arrange informed review when the issue exceeds your present expertise. These actions develop capability more directly than collecting another set of impressive answers.

Audy can now describe why a learner task, an activity, and a check need to agree in her example. To reuse that understanding in unfamiliar situations, she must go deeper than the surface wording of the brief. Chapter 5 examines how to unpack the concepts and assumptions that make such judgments possible.

References

Bastani, H., Bastani, O., Sungu, A., Ge, H., Kabakcı, Ö., & Mariman, R. (2025). Generative AI without guardrails can harm learning: Evidence from high school mathematics. Proceedings of the National Academy of Sciences, 122(26), Article e2422633122. https://doi.org/10.1073/pnas.2422633122

Roediger, H. L., III, & Karpicke, J. D. (2006). Test-enhanced learning: Taking memory tests improves long-term retention. Psychological Science, 17(3), 249–255. https://doi.org/10.1111/j.1467-9280.2006.01693.x