Appendix B — Prompt Library

Use these patterns with the ethics and data gate. Replace bracketed fields with approved material. A role instruction is not a privacy control, a source of professional authority or a guarantee of factual accuracy.

The examples refer to the fictional PDP-2026.1 pack. With no AI access, use its fixed responses and have a partner read the bounded role card. Keep the same outcome, sources and assessment criteria.

B.1 Before any prompt

Write down the capability students should demonstrate, permitted assistance, allowed source records and a stopping condition. Decide how feedback will be checked and what the learner will subsequently do without that assistance.

Use this common boundary where relevant:

This is a fictional teaching exercise.
Use only the supplied case records.
Separate recorded facts, inference and missing information.
Do not invent records, policy, quotations or observed outcomes.
Ask for a missing essential input instead of filling it in.
Do not request personal, medical or confidential information.
Your response is provisional material for educator review.

This instruction can reduce ambiguity but cannot enforce compliance. Inspect the response before showing it to learners. Do not include real sensitive data simply because the instruction says it is fictional.

B.2 Starting framework: CRAFT: specify a bounded task

CRAFT stands for Context, Role, Action, Format and Tone/Target. It organises a request; it does not certify the answer.

Context: fictional supplier task S1.
We need 200 units within five working days and have $4,000.
Use the attached S1 offers, with no additional assumptions.

Role: a drafting assistant, not the purchasing approver.
Action: compare full-quantity cost and stated capacity.
Format: a short calculation for each offer,
then one recommendation and one unresolved check.
Tone/Target: plain language for a junior procurement colleague.

Do not assume a cheaper offer can supply missing units.

Check: A costs $3,800, B’s 200-unit price would be $3,500 but exceeds its stated capacity by 40 units, and C costs $4,000. The recommendation must respect both price and capacity. Consult the actual S1 card for all task constraints.

No-AI route: calculate these figures and exchange recommendations for checking.

B.3 1. Reverse prompting: ask for essential inputs

I am designing a practice activity for [outcome].
Before proposing the activity, ask up to four questions
about the sources, permitted assistance, access needs
and independent task. Ask one at a time.
Do not request student identities or actual private records.
If I cannot answer, record the gap rather than inventing it.

Check: useful questions concern the proposed learning task. A request to upload a complete student file is not necessary merely because the model asked for it. Answer with fictional or approved information, or stop.

No-AI route: use the four categories as a planning checklist with a colleague.

B.4 2. Pros and cons: check rather than oppose

Compare offers A, B and C in S1 against the given
quantity, deadline and budget.
State one supported advantage and constraint for each.
Do not add an invented disadvantage for symmetry.
Wait while I check your arithmetic and recommendation.

Check: a learner may agree with the recommendation after checking it. Use the fixed offers on paper for the same comparison.

B.5 5. Debating: distinguish arguments from evidence

Using the supplied fictional teaching sources,
draft two competing recommendations on [decision].
For each, identify the supporting record, trade-off
and missing fact. Label this as a generated debate.
Do not invent citations or declare a consensus.
Ask me to justify which recommendation, if either, I retain.

Check: model-generated disagreement is not evidence of real stakeholder disagreement. On paper, write the strongest supported alternative to your first recommendation and explain your decision.

B.6 3. Stepwise tutoring: make the learner attempt

Use the supplied M1 marketing figures.
Ask me to calculate campaign A's click-through rate.
Wait for my calculation and explanation.
Check it against the supplied numerator and denominator.
If wrong, give a small hint and ask me to try again.
Then ask for B's rate and what comparison is justified.
Do not treat "next" as evidence that I understand.
Do not infer profit or causal impact from these records.
Stop after four learner attempts or when I say STOP.

Check: the rates are 4% and 5%; both campaigns have 20 recorded purchases. Different audiences and weeks prevent a clean causal conclusion. If the model praises a wrong denominator, correct the feedback.

No-AI route: use a worksheet with the calculations covered until the learner has attempted them.

B.7 Additional pattern: Socratic questioning: probe a decision

I recommend [decision] using [approved source IDs].
Ask one question about the weakest supported assumption.
Wait for my answer, then ask one question about authority
or a missing record. Do not imply I must reverse my view.
After two questions, summarise what remains unresolved.
Do not supply the independent assessment answer.

Check: questioning should expose an actual gap, not pressure learners into disagreement. A justified decision may remain unchanged.

No-AI route: partners take turns asking “Which record supports that?” and “What would change your decision?”

B.8 4. Role-play: a bounded exchange

Play Alex using only P1-D4.
I am the junior adviser whose authority is in P1-D5.
The exercise concerns the first response, not a final ruling.
Begin with Alex's stated concern about dependencies.
Do not add private history, new policy or unseen evidence.
If asked for an absent fact, say the exercise does not supply it.
Stop after six exchanges or immediately on STOP.
Do not score me, diagnose anyone or demand personal disclosure.

Check: review the dialogue against P1-D1–D5. A plausible invented fact changes the case and should not silently become evidence. Debrief the response separately using the educator rubric.

No-AI route: a partner reads the role card, or the learner responds to the fixed statement in the pack.

B.9 7. Generated perspectives: inspect alternatives

Using only InnovateCo I1-D1 to I1-D5 and its four role cards,
draft one concern each for Jamie, Priya, David and Sarah.
Label these as generated perspectives, not interviews.
For each, identify a supporting record and an unknown.
Do not add facts or treat agreement as corroboration.
Finish with two questions for a real decision-maker to resolve.

Check: four outputs from one model are not four independent witnesses or experts. Compare each assertion with the role cards. Keep unresolved matters open.

No-AI route: distribute the four cards among a group, or compare their fixed statements individually.

B.10 Additional pattern: critique and revision: test consequential claims

Here is response P1-A and the P1 records.
List up to three consequential claims in the response.
For each, give the exact source ID or say "unsupported".
Identify any decision beyond the junior adviser's authority.
Do not assign a mark or invent a missing policy.
Wait while I decide which changes to make.

Check: P1-A invents an April warning, presumes a performance plan, approves hours and asks for unnecessary family details. An AI critique that misses these is incomplete, even if its prose is fluent.

No-AI route: annotate P1-A directly and compare with the educator notes only after completing your own analysis.

B.11 6. Formative practice and feedback

Use after the learner has made an initial judgement:

Compare my draft with the five supplied practice criteria.
For each concern, quote the relevant phrase from my draft
and identify the case record or criterion involved.
Distinguish an error from a stylistic preference.
Do not award a grade, predict an official result
or assume all your suggestions should be accepted.
End with one question for my own revision decision.

Check: the author-created flawed feedback P1-F awards 90% to a response the illustrative educator calibration scores 4/20. Students should show why feedback is retained, revised or rejected. Do not teach automatic opposition.

B.12 Context handover

Draft a handover for the next practice session:
task and capability; source version and IDs;
confirmed facts; proposals not yet approved;
missing records; limits of authority; next question.
Keep it under 200 words.
Do not convert a suggestion into an agreed action.
I will check and approve it before reuse.

Check: compare the handover against the source records, not just the previous generated summary. Carry the approved source pack forward too. Context size and product memory do not ensure faithful recall.

B.13 A small content transformation

Turn the supplied M1 exercise into a self-contained HTML page.
Keep every figure unchanged.
Use headings, visible labels and ordinary text areas.
Include the task, source figures and three response fields:
claim, numerical check and justified next decision.
Use no external scripts, fonts, analytics or network calls.
Do not claim LMS submission, persistent saving
or accessibility certification.
Explain how an educator can test export and printing.

Check: compare a generated page with the supplied decision sheet. Verify calculations, keyboard use, readable print and the actual exported file. A useful static page still needs device and LMS testing before adoption.

B.14 Assessment design support

Here is my outcome, specimen task and marking guide.
Identify where assistance could produce a good submission
without showing the learner's independent capability.
Propose one changed task that tests the same outcome
under the stated permitted conditions.
List new access and workload risks.
Do not claim the redesign is cheat-proof or accredited.

Check: an educator must attempt the changed task, moderate specimen responses and review reasonable adjustments. A model’s confidence is not assurance. Follow the stress-test sequence.

B.15 Research and sensitive material

This library is not a research ethics protocol or a method for automated thematic analysis. Do not paste participant transcripts, student records or confidential workplace material into a tool on the basis of these examples. For an approved research use, agree the method, consent or other authority, data handling, access controls and analysis responsibilities first.

For locating reading, ask a tool for search terms, then retrieve and read the original works yourself. Do not treat generated citations, summaries or consensus among generated personas as verified literature.

B.16 Record a usable prompt

Keep a short teaching record: activity/version, intended capability, supplied source IDs, approved product and date if used, important observed failures, checked feedback and the next independent task. Record only what is necessary under the approved retention arrangement. A prompt without its sources and checks is not a reproducible activity.