Reviewed, classroom-ready AI prompt templates

Event notice for parents (slot-based)

The reusable notice template — fill five slots, review, send.

Template

School event notice for parents/guardians. Warm, respectful, ~{{WORD_CAP}} words, English and {{SECOND_LANGUAGE}} in parallel. Include: {{EVENT}}, {{DATE_DAY}}, {{TIME}}, {{WHAT_TO_BRING}}, contact: {{CONTACT_SLOT}}. No student names; no urgency pressure; nothing that assumes a two-parent household. Offer one shorter WhatsApp-style version too.

Review: Pre-send review: every date/day/time/amount checked against the source; no unauthorised commitments; no personal data; the sent version is the reviewed version.

Privacy: Never include student names, identifiable marks, health, family or community details. Describe situations and needs, never children.

Known failure modes: Invents the day-name for the date — check date/day pairing against a calendar · Adds 'refreshments will be provided' style commitments nobody made

Meeting agenda with timings and decision flags

Turn a topic list into a timed agenda with decisions-needed marked.

Template

A staff-meeting agenda from these topics: {{TOPIC_LIST}}. Total time {{TOTAL_MINUTES}} minutes. Assign realistic minutes per topic, order for energy (decisions early, information later), mark each item [DECISION] or [INFO], and add a 5-minute AOB slot. No names attached to items; a one-line purpose per item.

Review: Pre-send review: every date/day/time/amount checked against the source; no unauthorised commitments; no personal data; the sent version is the reviewed version.

Privacy: Never include student names, identifiable marks, health, family or community details. Describe situations and needs, never children.

Known failure modes: Uniform minutes per item regardless of weight — adjust before circulating

Routine checklist (lab day, exam day, event day)

A before/during/after checklist for any recurring operational day.

Template

A {{DAY_TYPE}} checklist for a teacher: BEFORE (preparation items), DURING (monitoring items), AFTER (closure items) — one line each, checkbox format, max {{ITEM_CAP}} items total, ordered by time. Specific to: {{CONTEXT_DETAILS}}. No generic filler ('be prepared') — every item is a checkable action.

Review: Pre-send review: every date/day/time/amount checked against the source; no unauthorised commitments; no personal data; the sent version is the reviewed version.

Privacy: Never include student names, identifiable marks, health, family or community details. Describe situations and needs, never children.

Known failure modes: Generic filler items — 'checkable action' is the bar

Blueprint-driven question paper

A full test built row-by-row from an outcome/level/marks blueprint.

Template

{{BOARD}} Class {{CLASS}} {{SUBJECT}}, chapter(s) {{CHAPTERS}}. Build to this blueprint: {{BLUEPRINT_ROWS}} (each row: outcome — item count — type — difficulty — marks). Plus one misconception item targeting {{KNOWN_MISCONCEPTION}}. Answer key with one-line reasoning per item, items labelled by outcome and cognitive level. Syllabus notation only; distractors must be plausible errors, never jokes; I verify every key. Flag your least-confident item.

Review: Solve every item yourself before trusting the key; check each item maps to a taught outcome; run the six-defect audit (ambiguity, keys, difficulty honesty, context bias, distractors, curriculum match).

Privacy: Never include student names, identifiable marks, health, family or community details. Describe situations and needs, never children.

Known failure modes: Ignores the blueprint totals — recount marks before use · Key errors — solve every item

Feedback drafts from anonymised patterns

Encouraging, specific feedback language from performance patterns — never pasted student work.

Template

Task: {{TASK}}, rubric criteria: {{CRITERIA_LIST}}. For each of these anonymised performance patterns, draft feedback (two strengths, one next step, encouraging tone, ≤60 words each): {{PATTERN_LIST}}. Address the work, not the person; next steps must be actionable this week. I will personalise offline.

Review: Feedback drafts are language, not judgement — the level/mark is read from the work by you.

Privacy: Describe performance patterns only. Never paste identifiable student work into a public tool. Never include student names, identifiable marks, health, family or community details. Describe situations and needs, never children.

Known failure modes: Generic next steps ('work harder') — actionable-this-week is the bar

MCQs with error-based distractors

MCQs where every wrong option encodes a real, named student error.

Template

Class {{CLASS}} {{SUBJECT}}, outcome: {{OUTCOME}}. {{ITEM_COUNT}} MCQs where each wrong option encodes a REAL error a student makes — after each item, one line naming the error each distractor represents. No joke options, no 'all of the above'. Answer key; I verify. Difficulty: {{DIFFICULTY_MIX}}.

Review: Solve every item yourself before trusting the key; check each item maps to a taught outcome; run the six-defect audit (ambiguity, keys, difficulty honesty, context bias, distractors, curriculum match).

Privacy: Never include student names, identifiable marks, health, family or community details. Describe situations and needs, never children.

Known failure modes: Distractors drift into implausible options — the named-error line exposes this

Project rubric with observable descriptors

A rubric two markers would apply identically, plus a student-language version.

Template

Class {{CLASS}} project: {{PROJECT_DESCRIPTION}}. A rubric with {{CRITERIA_COUNT}} criteria ({{CRITERIA_LIST}}), {{LEVEL_COUNT}} levels each, and OBSERVABLE descriptors per cell — something a reader could point to in the work, no adjectives without evidence. Plus a student-language version of the same rubric. Flag any descriptor two markers could still read differently.

Review: Test-score one imaginary strong and one weak submission: if you hesitate on a cell, iterate that descriptor.

Privacy: Never include student names, identifiable marks, health, family or community details. Describe situations and needs, never children.

Known failure modes: Adjective descriptors ('good understanding') — the observability rule catches them

Inquiry stimulus pack (scenario + anomaly + open questions)

Start an investigation without resolving it — conclusions forbidden.

Template

Class {{CLASS}} {{SUBJECT}}, topic {{TOPIC}}. Produce an inquiry stimulus pack: a one-page fictional scenario, a data table with ONE planted anomaly ({{ANOMALY_HINT}}), and three genuinely open investigation questions. NO conclusions, no explanation of the anomaly, no model answers anywhere in the output. Fictional setting only. Tell me where you were most tempted to explain too much.

Review: Check the output contains NO conclusions or model answers where students must think; fictionalise any real-community setting; verify the stops-here line held. Check the anomaly is discoverable by a student at this class level.

Privacy: Never include student names, identifiable marks, health, family or community details. Describe situations and needs, never children.

Known failure modes: Answers leak into the questions — audit with the I-slot confession · Anomaly too subtle for the class level

Socratic question ladder (questions only)

Five questions from observation to principle, no answers included.

Template

Class {{CLASS}} {{SUBJECT}}, concept: {{CONCEPT}}. A five-question Socratic ladder from concrete observation to underlying principle: question 1 asks what students can directly observe/recall; each later question requires the previous answer; question 5 reaches the principle. QUESTIONS ONLY — no answers, no hints in brackets. Plus one line for me: where students most often get stuck on this ladder.

Review: Check the output contains NO conclusions or model answers where students must think; fictionalise any real-community setting; verify the stops-here line held.

Privacy: Never include student names, identifiable marks, health, family or community details. Describe situations and needs, never children.

Known failure modes: Later questions answerable without the earlier ones — the chain must be real

Accessibility-first format of a handout

The same content rebuilt with named accessibility features.

Template

{{BOARD}} Class {{CLASS}} {{SUBJECT}}, handout: {{HANDOUT_TOPIC}}; same content, two formats. Format A standard; Format B accessibility-first: sentences under {{WORD_LIMIT}} words, one instruction per line, numbered sections, key terms bolded once with margin glosses, an alt-text line for every image carrying its teaching point, tables with a single header row and a one-line summary. Conceptual level constant — simplify syntax, not the content. List the accessibility choices you made so I can verify each.

Review: Verify each named feature actually appears — AI sometimes acknowledges a feature without applying it. Offer Format B to everyone, unlabelled.

Privacy: Never include student names, identifiable marks, health, family or community details. Describe situations and needs, never children.

Known failure modes: Concept watered down instead of syntax — the golden constraint must be stated · Alt text describes appearance, not the teaching point

Three same-goal tiers of one worksheet

Support, standard and extension versions that stay discussable together.

Template

{{BOARD}} Class {{CLASS}} {{SUBJECT}}, worksheet on {{TOPIC}}; same outcome for every version: {{OUTCOME}}. Three same-goal versions: one with a word bank and sentence starters, one standard, one with an extension that DEEPENS ({{EXTENSION_DIRECTION}}). Same core questions in all three so the class discusses together; no version labelled by ability anywhere on the page. Tell me which version you are least confident pitches right.

Review: Check the goal survived in every version (same core questions); read all tiers aloud for equal-respect tone; no ability labels on any student-facing page.

Privacy: Never include student names, identifiable marks, health, family or community details. Describe situations and needs, never children.

Known failure modes: Support tier quietly loses the analysis question — check the goal survived · Extension adds length, not depth

Five-minute spiral recap of older material

Daily spaced review: three quick items from past chapters to open any lesson.

Template

Class {{CLASS}} {{SUBJECT}}. A 5-minute opening recap of three items: one from last lesson, one from {{WEEKS_AGO}} weeks ago ({{OLDER_TOPIC}}), one from earlier in the year ({{OLDEST_TOPIC}}). Oral or slate-based, no printing. Include the expected answers. Items must be recall-or-one-step — this is retrieval, not a test.

Review: Verify the answers; keep it genuinely under five minutes.

Privacy: Never include student names, identifiable marks, health, family or community details. Describe situations and needs, never children.

Known failure modes: Items too hard for a warm-up — recall-or-one-step is the bar

Bilingual concept explanation (parallel columns)

The same explanation in two languages carrying identical facts.

Template

{{BOARD}} Class {{CLASS}} {{SUBJECT}}, concept: {{CONCEPT}}. A bilingual explanation, ~{{WORDS_PER_LANGUAGE}} words per language in parallel columns (English / {{SECOND_LANGUAGE}}), SAME facts in both — nothing extra in either. Technical terms per this policy: {{TERMINOLOGY_POLICY}}. Everyday register both sides. Then list every terminology decision you made in a table.

Review: Run the three parity checks: back-translation for meaning drift, terminology table row-by-row, read-aloud for register. High-stakes bilingual content needs a human reviewer of both languages before use.

Privacy: Never include student names, identifiable marks, health, family or community details. Describe situations and needs, never children.

Known failure modes: One column gains a fact the other lacks — the same-facts rule is checkable line by line

Parallel bilingual worksheet (same questions, same spaces)

One worksheet both language groups complete identically.

Template

{{BOARD}} Class {{CLASS}} {{SUBJECT}}, topic {{TOPIC}}. A one-page worksheet in parallel English–{{SECOND_LANGUAGE}}: every instruction line bilingual, questions numbered identically, same answer spaces. Terminology policy: {{TERMINOLOGY_POLICY}} (e.g. technical terms stay in English with bracketed {{SECOND_LANGUAGE}} on first use). Same reading level in both languages — everyday register, not literary. Provide the terminology table (term / English / {{SECOND_LANGUAGE}} / decision) for my review.

Review: Run the three parity checks: back-translation for meaning drift, terminology table row-by-row, read-aloud for register. High-stakes bilingual content needs a human reviewer of both languages before use.

Privacy: Never include student names, identifiable marks, health, family or community details. Describe situations and needs, never children.

Known failure modes: Second language arrives more formal than the English — state the register · Terminology decisions made silently — demand the table

Diagnose the wrong model behind an error pattern

Name candidate mental models producing a recurring wrong answer.

Template

Class {{CLASS}} {{SUBJECT}}. Frequent anonymised answer pattern: {{ERROR_PATTERN}}. Answer three things: (1) what underlying model(s) produce EXACTLY these answers, (2) the standard name for this misconception if one exists, (3) three diagnostic questions where each wrong option corresponds to a DIFFERENT candidate model — with an option-to-model map after each. Also say whether this is a true misconception or a normal developmental stage.

Review: Verify the counterexample and every answer key against the textbook — a wrong confrontation entrenches the misconception it was meant to break. Verify the diagnostic option-model maps yourself.

Privacy: Never include student names, identifiable marks, health, family or community details. Describe situations and needs, never children.

Known failure modes: Options that fit two models — each must map to exactly one · Skips the developmental-stage question — it changes everything

Four-part remediation pack (confront–rebuild–model–practise)

The complete targeted intervention for one named misconception.

Template

{{BOARD}} Class {{CLASS}} {{SUBJECT}}. Named wrong model: {{WRONG_MODEL}}. Produce four parts: (1) a CONFRONTATION — a counterexample where this model visibly fails; (2) a REBUILD — the correct model explained ≤{{REBUILD_WORDS}} words with one analogy at class level; (3) one worked example with the thinking said aloud; (4) {{ITEM_COUNT}} scaffolded practice items ordered from the confrontation to transfer, deliberately mixing items where the old rule coincidentally works with items where it fails. Answer key with reasoning; I verify key AND confrontation. Predict which practice item students will most likely still miss.

Review: Verify the counterexample and every answer key against the textbook — a wrong confrontation entrenches the misconception it was meant to break.

Privacy: Never include student names, identifiable marks, health, family or community details. Describe situations and needs, never children.

Known failure modes: Practice items all have the new shape — mixing is what forces model choice · Confrontation too subtle to land — it must fail visibly

Extract every citation for manual lookup

A lookup-ready list of all references in a document — fakes look perfect.

Template

Extract every citation, reference, statistic and named source from this content into a lookup table: source as stated / the claim it supports / where I could check it (textbook, official site, library). Do NOT tell me whether they are real — formatting cannot prove existence; I will look each one up. Flag vague attributions ('studies show', 'experts say') separately as unverifiable-as-stated. Content: {{PASTE_CONTENT}}

Review: Verify against real sources, never the AI itself; look up every citation; the approve/revise/reject verdict is recorded and yours. A fake reference is worse than none: look up or leave out.

Privacy: Never include student names, identifiable marks, health, family or community details. Describe situations and needs, never children.

Known failure modes: The model asserts citations are real — the do-not-tell-me instruction keeps the judgement offline

Run the eight-step verification tree on content

Structured claim/source/age/bias/objective review of any AI output.

Template

Run a verification review of this content for Class {{CLASS}} {{SUBJECT}} (objective: {{OBJECTIVE}}): (1) list every FACTUAL CLAIM it makes, (2) for each claim, state how I could verify it against the textbook or an official source (do NOT verify by your own confidence), (3) list every citation/source it names — I will look each up, (4) age-appropriateness concerns, (5) bias/exclusion concerns (who appears, who is missing), (6) does it meet the objective? End with your uncertainty list: the claims you are least sure of. Content: {{PASTE_CONTENT}}

Review: Verify against real sources, never the AI itself; look up every citation; the approve/revise/reject verdict is recorded and yours. The tree assists; the lookups and the verdict are yours.

Privacy: Never include student names, identifiable marks, health, family or community details. Describe situations and needs, never children.

Known failure modes: The AI 'verifies' its own claims confidently — the do-not-verify-by-confidence instruction blocks this framing, but the lookups remain yours

Score a prompt on the 10-point SATHI scorecard

Diagnose a weak prompt before regenerating — every flaw maps to a slot.

Template

Score this prompt on the SATHI scorecard — two points per slot (present? specific?), 10 total: S situation/subject, A audience/ability, T task/format, H human checks/boundaries, I iterate/improve. For each slot scoring under 2: name the missing element and write the one-line addition that fixes it. Then diagnose the overall disease if any: ambiguity / missing context / overconstraint. Prompt: {{PASTE_PROMPT}}

Review: Verify against real sources, never the AI itself; look up every citation; the approve/revise/reject verdict is recorded and yours.

Privacy: Never include student names, identifiable marks, health, family or community details. Describe situations and needs, never children.

Known failure modes: Generous scoring — ask for the fix lines regardless; they are the useful output