Prompt Engineering for Educators: A Case Study

The Problem Educators Faced

A high school history teacher named Jordan tried using AI to generate a differentiated reading comprehension worksheet for a class with three genuinely different reading levels. The output looked fine at a glance — clean formatting, reasonable questions — until she actually handed it out. The "modified" version for her lower-level readers used the exact same vocabulary as the standard version, just with shorter sentences. It hadn't actually simplified the content, just the syntax. Three students who needed real scaffolding got a worksheet that looked easier but wasn't.

That's the failure that pushed her to actually learn prompt engineering for educators instead of treating AI tools as a vending machine for worksheets. The underlying issue wasn't the AI being bad at its job. It was that her prompt asked for "a modified version" without defining what modification actually meant for her specific students.

⚠️ Common mistake: Assuming "simplify this" or "make it easier" gives the AI enough information to differentiate meaningfully. Real differentiation needs specifics: vocabulary level, sentence complexity, scaffolding type, and what stays conceptually the same versus what actually changes.

The Wrong Approach

Jordan's original prompt looked like this:

Create a reading comprehension worksheet about the causes of World War I, with a modified version for struggling readers.

This fails because "struggling readers" isn't a specification, it's a label. Struggling with vocabulary is a different problem than struggling with working memory or struggling with abstract concepts, and each needs a genuinely different kind of support. The AI defaulted to the most common interpretation — shorter sentences — because that's the statistically average thing "modified for struggling readers" tends to mean across generic training data.

What this does: produces a worksheet that looks differentiated on the surface without addressing the actual barrier a specific group of students faces, which means the modification does nothing for the students who actually needed it.

The Correct Prompt

Here's the rebuilt version Jordan uses now, with actual specifics about what needs to change and what needs to stay the same:

Create a reading comprehension worksheet about the causes of World War I for 10th grade students.

Standard version: grade-level vocabulary, 5 short-answer questions requiring textual evidence.

Modified version for students who struggle with academic vocabulary (not comprehension of concepts):
- Keep the same 5 questions and same conceptual content
- Replace academic vocabulary with everyday synonyms, but keep key historical terms (e.g. "alliance," "assassination") and define them in a glossary box
- Add one sentence starter per question to support written response structure

Both versions should take roughly the same time to complete.

What this does: separates the actual barrier (vocabulary, not concept difficulty) from the surface-level fix (shorter sentences), and the "keep key historical terms" instruction prevents the AI from oversimplifying content that actually needs to stay intact for the lesson objective to hold.

⚡ Pro tip: Specify explicitly what should NOT change between a standard and modified version. This stops the AI from accidentally removing content that's actually essential to the learning objective while trying to be helpful.

⚡ Pro tip: If you teach the same unit across multiple sections with different needs, build the "what stays the same / what changes" specification once and treat it as a reusable template, adjusting only the specific barrier named for each section's students.

Results and What Changed

The rebuilt worksheet took the three students who'd struggled with the original version from an average of 40% correct to 78% correct on the same underlying content — same concepts, same rigor, just an actual scaffold instead of a cosmetic one. Jordan says the biggest shift wasn't the AI getting better, it was her getting specific about what "struggling" meant for those particular students instead of using it as a catch-all label she'd been reaching for out of habit.

A middle school science teacher applied the same specificity principle to lesson plan generation after hearing about Jordan's experience at a department meeting. Instead of asking for "an engaging lesson on photosynthesis," she now specifies the exact misconception she wants addressed: "many students think plants get their mass from soil rather than air — build the lesson around correcting this specific misconception." The resulting lesson plans consistently hit the actual learning gap instead of a generic overview of the topic.

An instructional coach supporting a district-wide rollout of AI tools for lesson planning found that teachers who specified a concrete misconception or skill gap in their prompts produced measurably more targeted materials than teachers who described only the topic. That's a pattern worth remembering: precision about the actual gap beats precision about the subject matter every time. She now opens every training session with Jordan's worksheet story, because it makes the abstract idea of "specificity" concrete in a way that a slide full of bullet points never quite manages.

There's a broader point buried in all three of these examples: the AI wasn't the variable that changed between the failed attempt and the successful one. The teacher's own clarity about the actual problem was. That's a slightly uncomfortable thing to sit with if you were hoping better prompting techniques would substitute for knowing your students well, but it's also good news — the skill that makes these tools work well is exactly the skill good teachers already have. It just needs to get written down instead of staying implicit.

⚡ Pro tip: When generating any assessment or worksheet, ask the AI to explain its own scaffolding choices in a short note at the bottom. This gives you a fast way to catch a mismatch, like Jordan's shortened-sentences issue, before you print 30 copies and hand them out to a room full of students.

⚡ Pro tip: For differentiation specifically, describe the barrier (vocabulary, working memory, abstract reasoning, motor skills for handwriting) rather than a vague label like "struggling" or "advanced." The barrier determines the actual fix.

How to Apply This to Your Situation

Start by picking one worksheet, quiz, or lesson plan you generate regularly and rebuild the prompt with real specifics: grade level, the actual misconception or skill gap you're targeting, and — if you're differentiating — the precise barrier each modified version needs to address, not a generic label.

Test the modified version yourself before handing it to students, the way Jordan wishes she had the first time. A two-minute read-through catches the "same vocabulary, shorter sentences" problem before it reaches a kid who needed real support and didn't get it. This is a small habit, but it's the single most valuable check in this entire workflow, because it's the last point where a mismatch is cheap to fix instead of costly.

⚠️ Common mistake: Building one modified version and assuming it works for every student with that label. A student struggling with vocabulary and a student struggling with sustained attention need genuinely different supports, even if both get called "needs modification" on paper. If you have students with meaningfully different barriers in the same class, you likely need more than one modified version, not one catch-all fix.

Next Steps

Build a small library of your most common lesson types — the ones you generate weekly — with the specificity fixes baked into the prompt template itself, so you're not reconstructing "what does struggling actually mean here" from scratch every time you sit down to plan.

Try this exercise with your department: have each teacher bring one worksheet or lesson plan they generated recently, and ask them to name the specific barrier their "modified" version was supposed to address. If they can't name it precisely, that's the same gap Jordan found in her own worksheet — a label standing in for an actual specification. This exercise tends to surface the pattern faster than reading about it, because it's uncomfortable in a useful way to realize how often "struggling" or "advanced" gets used without a concrete meaning behind it.

It's also worth building this habit into any AI-assisted IEP accommodation drafting, though with an important caveat: those documents carry legal weight, and drafts should always go through the same review process as any other IEP documentation, not skip it because the draft looked polished. The specificity principle still applies — naming the actual accommodation need rather than a vague category — but the review step stays non-negotiable.

A district technology coordinator who trained teachers across six schools on this approach noticed a pattern worth sharing: teachers who'd used AI tools longest weren't necessarily the ones getting the best results. The teachers getting the best results were the ones who'd learned to notice when a label was doing the work a real specification should be doing — a skill that has almost nothing to do with how long you've used the tool and everything to do with how precisely you think about your students' actual needs before you start typing.

Once you've got a handful of these working reliably, save them somewhere versioned rather than in scattered chat histories you'll never find again — PromptABCD works well here, letting you keep a shared library of prompt templates other teachers on your team can adapt for their own classes instead of everyone rediscovering the same lesson about vague labels the hard way.