The IMPACTS Framework: A Complete Guide to Teacher AI Use

Why a framework is needed at all

In sessions with teachers across NCR schools, one pattern repeats. The gap is never technical skill.

Teachers learn new systems quickly. Learning new systems is most of what the job has become over the past fifteen years.

The gap is that nobody gave them a method for applying the professional judgment they already have. So they do the one step everybody knows, which is to type a request, and they skip everything that should happen before it and everything that should happen after.

Every prompting guide available online addresses the wording of that request. Give the AI a role, state your task, add your constraints. All reasonable, and all operating on the step that carries the least weight.

The seven steps

I is for Identify. Define what you actually need, and why, before opening anything. Not "create a worksheet" but the real outcome: what must this reveal, what cognitive demand should it place, what constraints matter, what would make it fail. This step also includes the ethical check: no student identifiable information enters an AI tool.

M is for Mine. Ask AI what you have not thought of, before asking it to produce anything. What misconceptions do students hold about this concept at this age? What question formats expose flawed reasoning rather than rewarding memorisation? What changes for learners who need support, and for those ready for extension? This is the step that produces the largest single improvement in output quality.

P is for Prompt. Turn that thinking into an instruction. Purpose, requirements, format, differentiation, what to avoid. Every line traces back to a decision made in Identify or Mine. This is a brief, not a request.

A is for Assess. Judge the output against criteria set in advance. Does it meet the defined purpose? Does it probe reasoning rather than recall? Is it factually correct? Is the language age appropriate? Is it realistic for your actual classroom?

C is for Calibrate. Correct with precision. Not "make it better", which produces a different draft of similar quality, but numbered, specific instructions: convert question three to explanation based, add a reasoning follow up, test the evaporation misconception directly.

T is for Tune. Decide when it is genuinely ready. This prevents both over-iteration, which wastes time, and under-iteration, which puts substandard work in front of children. The question is simply whether you would confidently use this tomorrow.

S is for Systematise. Stop rebuilding the same thing every term. Separate what is permanent, your grade, rubric, tone, format and standards, from what is occasional, this student, this topic, this week. Install the permanent layer once so future work is reviewed rather than rebuilt.

Why six of the seven sit outside the prompt

Count where the steps happen. One of them, Prompt, is the act of typing. The other six are thinking before, and judgment after.

That distribution is the central finding, not a design preference. The quality of what AI produces for a teacher is decided almost entirely by the clarity of intent brought to it and the rigour of evaluation applied to it.

This is why two teachers using the same tool, on the same task, in the same fifteen minutes, produce work of very different quality. And it is why buying a better tool rarely solves the problem a school thinks it is solving.

The three modes

Each step belongs to one of three modes.

AI as thinking partner: Identify and Mine. Used to expand what you consider before anything is produced.

AI as execution partner: Prompt, Assess, Calibrate, Tune. Used to produce and refine to a standard you set.

AI as standing capability: Systematise. Used to retain what you decided, permanently.

Most teachers only ever operate in the second mode. That is the whole problem, stated as compactly as possible.

Why this framework favours experienced teachers

Every one of the seven steps is a point where professional judgment enters the process. The framework is not a replacement for expertise. It is a structure for applying it.

That matters because AI output arrives at a fairly constant quality. What varies enormously is what happens next, when someone has to look at it and decide whether it is any good.

That decision requires knowing what good looks like, for this topic, this age group, this class, this term. A teacher of twenty years makes those judgments in seconds and often cannot explain how. A teacher who is quick with software but new to the classroom cannot make them at all.

In 1986, Lee Shulman named this knowledge. He called it pedagogical content knowledge, and described it as including "the conceptions and preconceptions that students of different ages and backgrounds bring with them to the learning of those most frequently taught topics."

That is precisely what AI does not have, cannot acquire from training data, and needs supplied. It is generated by presence in a classroom, and only by presence.

Which means the ceiling on what AI can give a teacher is not her technical skill. It is her professional judgment. Experienced teachers hold the scarce input.

This is consistent with the national picture. CENTA's 2025 survey found AI adoption highest among teachers with more than three years of experience, at around 75 percent. The people with the most classroom judgment are already the heaviest users. The full data is covered in what the data actually says about AI use by teachers in India.

What the framework is not

It is not a prompt template. Templates address the one step that matters least.

It is not a tool recommendation. The framework works identically in Claude, ChatGPT, Gemini or Copilot, and will continue to work when those are replaced.

It is not a way to remove teacher judgment from the process. It does the opposite: it inserts judgment at six defined points where most workflows currently have none.

How to start using it

If you adopt only one step, adopt Mine. Before asking AI to make anything, ask what your students typically get wrong about the topic. It takes about fifteen seconds and it changes the request that follows.

The second easiest is Calibrate. When output is wrong, resist "make it better" and give three numbered corrections instead.

Those two alone move most teachers from producing acceptable material to producing material they would defend.