AI detectors flagged 97.8% of human-written essays by non-native English speakers. No prompt here detects AI writing — they plan, differentiate, assess and give feedback that works.
Start with the thing that will cost a student their grade
Before any lesson planning, the finding every teacher using AI needs and almost nobody has been given.
In GPT detectors are biased against non-native English writers (Patterns, 2023), Weixin Liang, Mert Yuksekgonul, Yining Mao, Eric Wu and James Zou ran seven widely-used AI detectors against 91 human-written TOEFL essays and 88 essays by US eighth-graders.
On the eighth-graders' essays, the detectors were near-perfect.
On the TOEFL essays — all written by humans — they misclassified over half as AI-generated, with an average false positive rate of 61.22%. All seven detectors unanimously flagged 18 of the 91 as machine-written. And 89 of the 91 essays — 97.80% — were flagged by at least one detector.
The mechanism is visible in the data: the unanimously-flagged essays had significantly lower perplexity, meaning more predictable word choice. Detectors are not identifying machine authorship. They are identifying constrained vocabulary — and then the researchers demonstrated the whole thing collapses in both directions, because asking a model to enhance the word choices to sound more like that of a native speaker
substantially reduced the false positive rate. The same trick lets actual AI text walk straight through.
So a detector flags your English-language learners at scale, misses anyone who paraphrases, and produces a number that looks like evidence. The authors' own conclusion is a caution against their use in evaluative or educational settings, particularly when they may inadvertently penalize or exclude non-native English speakers.
No prompt in this pack detects AI writing, and none should. If you need to know whether a student can do something, assess them in conditions where the question is answerable — in class, orally, in draft stages, on their own prior work. That is an assessment design problem, and it is solvable. Accusation based on detector output is not.
What the pack is actually for
Fourteen prompts, and they cluster around the parts of teaching that consume evenings rather than the parts that require a teacher in a room.
None of them replaces professional judgement about your students, and all of them work from what you supply — your standards, your text, your rubric, your class. Where a prompt would need to know something about your students that you haven't told it, it should ask rather than assume. Everything here still needs your review before it reaches a class: these tools produce plausible material fluently, including plausible material that is wrong, which matters more in a classroom than almost anywhere else. See why AI models hallucinate.
Plan the arc before the lesson
Design a Unit Scope and Sequence comes first for a reason. A sequence of individually good lessons that don't build is the most common structural failure in planning, and it is invisible from inside any single lesson.
Write Measurable Learning Objectives is the prompt that makes everything downstream easier. An objective in student-outcome language — what they will be able to do — determines the assessment, the activities and the exit ticket. A fuzzy objective produces a lesson that cannot be checked, and you find out at the test.
Build a Complete Lesson Plan is the workhorse: objectives, hook, activities, checks for understanding, materials, in a real time slot. The time slot matters — most lesson plans fail because they contain forty-five minutes of material for a thirty-minute period.
Design a Project or Hands-On Activity covers work where the deliverable is the assessment. Worth noting that project work, done in stages with visible drafts, is also one of the more robust answers to the authorship problem above — not as surveillance, but because you saw the thinking happen.
Teach the thing
Explain a Concept Several Ways for Teaching is, for my money, the highest-value prompt on this page. Your best explanation is the one that worked for you, and it will fail for some proportion of any class. Having four genuinely different framings ready — an analogy, a concrete example, a visual, a formal statement — is the difference between reteaching and repeating.
Differentiate a Lesson for Varied Learners adapts an existing task up and down. This is the work that is genuinely, structurally impossible to do well for thirty students in a prep period, and where the time saved is real rather than notional.
Write Discussion Questions for a Text or Topic produces layered questions — comprehension through analysis to open debate — so the conversation has somewhere to go after the first confident answer.
Check whether it landed
Create Exit Tickets and Quick Formative Checks is small and disproportionately useful. A check that takes two minutes to answer and thirty seconds to scan tells you tomorrow's lesson before you plan it.
Generate a Quiz or Test With an Answer Key and Make a Practice Worksheet both produce the answer key alongside the questions — and both need checking against it. A wrong answer key teaches the wrong thing with the full authority of an answer key, which is a worse outcome than no worksheet.
Create a Grading Rubric is where consistency comes from. Written criteria and specific level descriptors are what make two pieces of work graded a week apart comparable, and what make a grade explicable to a student or a parent who asks.
Feedback, where the evidence is genuinely surprising
Kluger and DeNisi's meta-analysis in Psychological Bulletin (1996) pooled 607 effect sizes across 23,663 observations. Feedback interventions improved performance on average (d = .41) — and over one third of them made performance worse.
Their explanation is the part teachers can act on. Feedback effectiveness decreases as attention moves up the hierarchy closer to the self and away from the task.
Feedback that directs attention to the work improves the work. Feedback that directs attention to the person — praise, comparison to peers, a bare grade, anything that invites what does this say about me
— is where the negative third lives.
Give Constructive Feedback on Student Work is built on that: name specific strengths in the work, and give one clear, doable next step. Not a general assessment of the student. The instinct to soften criticism with personal encouragement is kind and, on this evidence, frequently counterproductive.
The learning pack is the student-facing counterpart, and it is worth reading if you are advising students on how to use these tools — particularly on the distinction between getting an answer and learning something, which is the whole problem in miniature.
The rest of the job
Draft a Message to a Parent or Guardian handles a genuinely hard writing task: factual, kind and constructive at once, about a child, to someone who loves them. Getting the tone wrong costs a relationship you need for the rest of the year.
Plan for a Substitute Teacher turns your lesson into something self-contained that a stranger can actually run — the task always done at the worst possible moment, usually at 6am while ill.
Where this stops
These prompts do not know your students, your school's policies, or your jurisdiction's requirements. They will produce material aligned to standards you paste in and cannot verify that you pasted the right ones.
Check every factual claim, every worked answer, and every answer key before it reaches a class. Nothing here should be used to detect AI use, assign a grade without your judgement, or make a determination about a student's academic integrity. Anything touching special educational needs, safeguarding, disability accommodation, or a formal disciplinary process belongs with the qualified people and processes your institution has for exactly that — not with a prompt, and not with a detector.
Sources
- Weixin Liang, Mert Yuksekgonul, Yining Mao, Eric Wu and James Zou, GPT detectors are biased against non-native English writers, Patterns (Cell Press), 2023 — seven detectors against 91 human-written TOEFL essays and 88 US eighth-grade essays; near-perfect accuracy on the latter, a 61.22% average false positive rate on the former, 97.80% of TOEFL essays flagged by at least one detector, and the bias traced to lower perplexity rather than machine authorship
- Avraham N. Kluger and Angelo DeNisi, The Effects of Feedback Interventions on Performance: A Historical Review, a Meta-Analysis, and a Preliminary Feedback Intervention Theory, Psychological Bulletin 119(2), 1996, 254–284 — 607 effect sizes, 23,663 observations; average d = .41, but over one third of feedback interventions decreased performance, with effectiveness falling as attention moves from the task toward the self