Prompt Engineering for Teachers: Better Lesson Plans, Differentiation, and Feedback

Start with the learning goal, not the tool

Most disappointing AI lesson plans fail before the first prompt is typed. The teacher asks for “a lesson on fractions” and gets a generic worksheet that could fit any class, any age, and any curriculum. The problem is not the model. It is the missing context: who the students are, what they already know, what success looks like by the end of the lesson, and how much time you actually have.

Write the goal first, in one sentence, before you open the AI tool. For example: “By the end of 45 minutes, Year 5 students can add fractions with different denominators and explain their working in words.” That sentence does more work than any clever prompt trick. It tells the model the age group, the skill, the time limit, and the evidence you will accept. If you cannot write that sentence, stop and clarify the lesson. No prompt can rescue a vague objective.

Next, add three facts the model cannot guess: what students learned last week, the common misconception you expect, and the materials you already have. A prompt that says “students confused numerators and denominators last Tuesday, and we have fraction strips” will produce something usable. A prompt without those facts will produce something pretty and wrong for your class.

A reusable lesson-plan prompt pattern

Use the same structure every time, and change only the details. The pattern has five parts — role, students, goal, constraints, and output format. Keep it in a note you can copy and reuse.

Notice what the pattern forces you to decide: the output format. “Give me a lesson plan” invites a wall of text. “Give me a timed plan with a teacher script, three worked examples, and one exit question” gives you something you can teach from. Ask for timings in minutes, not vague phases like “introduction” and “main activity.” Ask for the exact questions you will ask, not a description of the discussion.

Then iterate once, on purpose. Read the first draft as a sceptic: Is the first example too hard? Does the timing add up to 45 minutes? Would the weakest reader in the class understand the instructions? Send one follow-up that fixes the weakest part, such as “Make example 1 use denominators 2 and 4 only, and rewrite the instructions at a Year 3 reading level.” One focused revision beats five new prompts from scratch. For a deeper method for improving prompts over time, see the guide to iterative refinement.

Differentiation: one lesson, three entry points

Differentiation is where AI genuinely saves time, because the hard part is not inventing three versions of a task — it is keeping all three versions aimed at the same goal. Ask the model to vary the support and the complexity, never the learning objective. Every student should leave able to do the same core skill, even if the route there differs.

Put the guardrail in the prompt itself: “All three versions must practise adding fractions with different denominators. Do not turn the support version into a different topic, and do not let the stretch version skip the core skill.” Then check the output against that rule before you print anything. If a version drifts — for example, the support group ends up only shading pictures — rewrite that version by hand or ask for a targeted fix. The model is a drafting assistant, not the teacher. You remain the person who decides whether the work matches the goal.

Feedback prompts that teach, not just correct

AI feedback is most useful when it shows the student the next step, not merely the right answer. A prompt that says “mark this work” usually returns a score and a list of errors. A prompt that says “point out one strength, one specific error, and the exact next step, in language a 10-year-old can act on” returns something a student can use without you translating it.

Always paste the success criteria into the prompt alongside the student work. Without them, the model invents its own standard, and that standard may be stricter, looser, or simply different from yours. With them, the feedback stays anchored to what you taught. For example: “Success criteria: correct common denominator, correct numerator, answer simplified. Comment only on these three things.”

A feedback pattern worth copying

• Name the strength first, with evidence: quote the line that worked, so the praise is specific.
• Name one error, not five: pick the error that blocks the next step and explain it in one sentence.
• Give one action: “Try question 3 again, using 12 as the common denominator.”
• End with a question: “Why did 12 work when 6 did not?” so the student has to think, not copy.

Never paste identifiable student work into a public AI tool, and never let AI feedback be the final word on a grade. Use it as a first draft you review, especially for students who are struggling: the tone matters as much as the accuracy, and a blunt machine comment can undo a week of confidence. Read every comment before it reaches a student. If you would not say it yourself, do not send it.

Guardrails: accuracy, privacy, and your own judgement

AI tools hallucinate confidently, including in subjects you teach every day. Dates, definitions, worked solutions, and quotations all need checking. Build a two-minute verification habit: check every factual claim in a generated lesson against your textbook or scheme of work, and rework every maths example yourself before you teach it. If you cannot verify it quickly, cut it. A shorter lesson you trust beats a longer one you hope is right.

Privacy is the other non-negotiable. Do not enter student names, photos, grades, or anything that could identify a child into an AI chat. Describe the class in general terms — “a Year 5 class, mixed ability” — and keep individual work anonymous. Check your school policy before using any new tool with student material, and prefer tools your school has approved. When in doubt, use the AI to prepare materials, not to process student data.

Finally, keep the decisions that belong to you. The model can suggest a seating plan, a grouping, or a grade boundary, but it does not know the child who had a bad morning, the parent conversation from last week, or the behaviour dynamic in row three. Use AI output as a proposal. Your professional judgement is the edit, not an optional extra.

A 20-minute weekly routine

Prompt engineering pays off when it becomes a routine, not a one-off experiment. Set aside 20 minutes once a week — ideally when you plan the following week — and run the same four steps for your hardest lesson.

1. Write the goal sentence and the three facts the model cannot guess.
2. Generate one lesson plan with the five-part pattern, then do one focused revision.
3. Ask for three entry points with the same objective, and check each against the goal.
4. Save what worked in a shared document with the prompt, the subject, and one line on what you changed.

That saved prompt library is the real asset. After a month, you will have a dozen prompts you trust, each already tested on your students. New teachers in your department can start from those instead of from a blank chat box. The playbook approach — a small set of proven prompts, reused and refined — is exactly what the prompt engineering playbook recommends for teams, and it works just as well for a single teacher. When you want to compare approaches side by side, run the same prompt with two different framings and keep the one whose output needed fewer corrections; the comparison habit is covered in few-shot versus zero-shot prompting.

Start with one lesson this week: one goal sentence, one revision, and one saved prompt.

Related reading

Build the habit behind this article with three companion guides: the prompt engineering playbook for a reusable team approach, iterative refinement for improving a prompt step by step, and few-shot versus zero-shot prompting for choosing the right amount of example in each prompt.

{{小节标题}}

段落正文。行内命令这样写 brew install ollama。

{{子标题，可选}}

• 要点：说明
• 普通条目

1. 步骤一
2. 步骤二

Limits and notes

收尾：适用边界、注意事项。