How to Get Better Answers from AI: A Practical Guide for Educators

The first response from an AI tool may be close enough to look useful and still be wrong for the work in front of you.

The format may be right while the level of difficulty is off. A rubric may sound polished but reward the wrong things. A set of questions may cover the topic without revealing what students understand. That is the point when many people either accept the draft too quickly or start over with a brand-new prompt.

There is another option. Diagnose what missed, change one instruction, and compare the new version with the first.

The skill here is reviewing the response.

Name the problem you can see

"Make it better" does not give the AI much to work with. Say what the response got wrong.

Perhaps the language is too advanced for the students who will use it. Maybe the response added requirements that were never part of the assignment. Maybe every rubric descriptor uses vague words such as "excellent" and "adequate" without telling a student what those words mean.

Point to the visible problem:

• "The criteria are appropriate, but the descriptors are too vague for students to use while revising."
• "The questions check recall. I need students to explain how the evidence supports the claim."
• "This version assumes students have access to materials that are not available in our classroom."

That kind of follow-up is useful because you are responding to an actual draft. You have something specific to accept, reject, or change.

Change one instruction at a time

When a response misses in several ways, it is tempting to send a long list of corrections. Sometimes that is necessary. If you are still learning what the tool can do with the task, one change at a time makes the result easier to judge.

Ask for a revision that addresses the most important problem. Keep the rest of the response stable. Then compare the versions.

Did the revision fix the issue you named? Did it create a new problem? Did the tool preserve the parts you wanted to keep?

You are learning about the response, but you are also learning how clearly you can describe the quality you need.

Show the shape of a useful answer

Examples help when a direction is difficult to express as a rule.

Suppose you are building a rubric for a short student presentation. You want the evidence criterion to describe observable work, not rely on labels such as "strong" or "weak." You could supply one model descriptor:

Then ask the AI to write the remaining levels with the same kind of specificity. The example does not decide the rubric for you. It shows the level of detail and the type of evidence you expect.

You can do the same with format. If you want a one-page table, provide the column labels. If you want feedback phrased as a question, include one example. If the audience needs plain language, show a sentence that has the right level of directness.

Use fictional or non-sensitive examples. You do not need to paste identifiable student work into a tool to show what useful feedback looks like.

Look for assumptions in the response

A polished response can hide the guesses underneath it.

After you receive a draft, ask:

The list may point to an unstated grade level, available technology, prior knowledge, or definition of quality. Check each item yourself. A generated list is another draft, not evidence of how the tool reached its answer.

Review those assumptions before requesting another version. Correct the ones that matter. Ignore the ones that do not.

This step does not make an AI response reliable by itself. It gives you another view of what needs checking.

Know when to stop revising

More prompting does not guarantee a better result.

If you are correcting the same issue repeatedly, rewriting most of the response yourself, or spending more effort explaining the task than completing it would require, stop. Keep the useful pieces if there are any. Finish the work another way.

The decision to stop is part of using AI well. A tool can be helpful for one stage of a task and unhelpful for the next.

A practice round

Use a tool your school permits and work with fictional or non-sensitive material.

Start with this request:

Read the response before asking for a revision. Choose one visible weakness and name it precisely. If the descriptors are vague, provide one model. If the tool added another criterion, tell it to remove that criterion while keeping the rest of the table stable.

Then compare the versions:

• What changed after your instruction?
• What stayed the same?
• Did the revision solve the problem you named?
• Is another pass worth the effort?

The prompt and examples in this issue are illustrative. They are not tool tests, guarantees, or claims about what any product will produce.

Keep student information, sensitive records, and private circumstances out of a practice prompt. Follow your school's guidance on approved tools and data use.

Put the review process to work

If you are choosing your first task, start with practical AI for teachers: where to start. To apply the same review process to classroom planning, read how to use AI for formative assessment. For a communication task, try the workflow for drafting your weekly parent newsletter with AI.