AI GUIDEUse an AI assistant to organize a few notes into a draft you can inspect. This is useful when you know the situation but have not decided how to present the next steps. You need a small set of notes and enough knowledge of the task to notice a mistake.
The important result is a plan whose facts you can trace back to the input. You still decide whether its suggestions make sense.
Start with material you can share
For your first attempt, use the fictional notes below. Do not paste a real student record, client conversation or internal incident report into an unfamiliar service. Replacing a name is not enough when the remaining details identify someone. If the task needs real information, first check your organization’s rules and the service’s handling of that account’s data. A provider’s policy can differ across consumer and commercial products; Anthropic’s commercial-data policy is one concrete example, not a policy for every AI tool.
Give the assistant a small clear job
Here are notes for a fictional adult English learner. They are an illustration, not a tested teaching recommendation.
N1. Goal: prepare a 30-minute conversation session about travel.
N2. The learner wants practice asking follow-up questions.
N3. Avoid homework requiring a paid app.
N4. The learner can study independently for 10 minutes on weekdays.
N5. The date of the next session is not confirmed.
Ask for one bounded transformation
Turn the notes into a draft session plan.
Use only the notes for facts about the learner and scheduling.
Separate supplied facts, proposed activities and open questions.
Make the session activities total 30 minutes.
Link each activity to the note it serves.
Keep independent study separate from the session.
Do not invent dates, achievements or personal details.
If information is missing, list a question rather than filling the gap.
Do not send messages, schedule anything or update records.
The instruction names the output and its constraints, a practice supported by Claude’s prompting documentation. It does not guarantee compliance. Read the response against the notes.
Inspect a proposed result
An editor-written example of an acceptable draft is:
* 5 minutes; Proposed activity: Discuss a travel situation chosen by the learner; Connection to the notes: Travel conversation, N1
* 10 minutes; Proposed activity: Practise asking follow-up questions about that situation; Connection to the notes: Follow-up questions, N2
* 10 minutes; Proposed activity: Role-play a travel conversation; Connection to the notes: N1 and N2
* 5 minutes; Proposed activity: Review two follow-up questions the learner wants to reuse; Connection to the notes: N2
Optional independent practice: on a weekday, spend up to 10 minutes writing a short travel exchange with follow-up questions, using paper or an existing free writing tool. This is a proposal based on N2–N4, not a new requirement or a claim of improvement.
Open question: when is the next session? N5 leaves this unresolved.
The example keeps facts separate from proposed activities. Its timings add up to 30 minutes. A teacher must still judge whether the language and activity difficulty suit the learner.
Look for a convincing mistake
Suppose a response instead says: “Next Tuesday, begin a paid speaking-app course. The learner has already mastered travel vocabulary.” It has invented a date, contradicted N3 and claimed an achievement absent from the notes. Better formatting would not fix it.
Check each factual statement against N1–N5. Mark suggestions as suggestions. Check the total time and any resource requirement. Look for a constraint that disappeared. NIST identifies confabulation and privacy as relevant generative-AI risks; its Generative AI Profile also calls for checking generated sources and citations. In this exercise, the numbered notes are the evidence you inspect.
Try a different task yourself
Use these fictional project notes:
P1. We need a one-page onboarding checklist draft.
P2. It must use the existing approved handbook.
P3. Its owner and deadline have not been agreed.
A proposed plan says: “Maya will publish the checklist on Friday.” Identify the unsupported decisions before asking AI to revise anything.
A successful correction leaves the owner and deadline unresolved, preserves the one-page limit, requires the handbook, and keeps the output a draft. If the handbook is not supplied, ask for it rather than inventing its contents. There is no single correct wording.
Adapt the method
A tennis coach can structure fictional observation notes; a teacher can prepare an outline; a learner can organize study time; a CTO can draft an experiment brief. In every case, supply the constraints, preserve unknowns, inspect the evidence and retain the human decision. The method organizes information; it does not validate sports, teaching or engineering judgments.
Next: A04, checking a summary against its original. If the process repeats, continue to A12, deciding what to automate. Both are planned lessons in this draft curriculum.
Sources
Read the original sources to explore the topic and check specific claims.