The teachers who get real value out of an AI assistant are the ones who already know what a good lesson looks like, because every judgment the tool cannot make falls back to them. AI skill is an extension of teaching skill. What is worth learning is less how to phrase a request and more how to decide what to ask for, what to check, and what not to give it at all.

This is about your own practice. Teaching students to use AI is a different job with a different set of decisions, and it lives at teaching AI literacy to students.

The skill is knowledge, verification, and refusal

Teachers who treat prompt-writing as the whole skill end up judging output by how confident it sounds, and confidence is the property AI produces most reliably. Prompting is a small procedural piece. The rest is subject knowledge, verification, ethical judgment, attention to privacy and bias, honesty about when AI was used, and the willingness to not use it at all.

Use AI like someone in the discipline would. A historian reads a claim and asks where it came from, who benefits from it, and what the record supports. A teacher of history working with AI runs the same reflexes over what the model hands back. Without them you have a fast writer with no accountability and no memory of your students.

Fluent use turns out to mean more reading. You read the draft against what you know, check the quotations, and notice the paragraph that would confuse the class that struggled with the same idea last week. The first version arrives sooner and you spend the saved minutes looking at it harder.

A 2026 systematic review of 28 empirical teacher–AI co-design studies found that teachers used AI mostly for ideation, lesson planning, and content production. Only a small portion of those studies showed AI working as a genuine partner in design. The recurring problems were generic output, misalignment with the actual curriculum, factual unreliability, bias, and opacity about how the system reached its answer. Efficiency crowded out the rest. The review’s recommendation: keep questioning, filtering, and adapting, and keep governance with the teacher.

UNESCO’s teacher framework warns that overreliance may weaken teacher agency and professional competency — worth sitting with before anything else here. TPACK revisited for generative AI, and the SEE framework’s safe-ethical-effective test, cut at the same problem differently.

What to hand the tool, and what stays with you

Before anything else you sort the work, and the sort follows one rule: AI suits tasks whose output is provisional, quickly checked, and not a high-stakes judgment.

Hand these to the tool freely. Alternative examples and nonexamples, quiz and question variants, first activity ideas, the misconceptions a topic tends to produce, draft discussion questions, first-pass differentiation. Also formatting, conversion into tables or organizers, draft feedback statements, several versions of the same directions, preliminary rubric language, family-communication drafts, and counterarguments to a lesson you have already designed. Each of these you can check in a minute, and a wrong one costs nothing.

Generate these, then review them substantially. Lesson sequencing, unit design, assessment items, exemplars. Reading-level adaptations, translations, culturally responsive examples, accommodations. Source packets, scoring assistance, and recommendations drawn from student work. These are worth generating and never worth accepting. The failure mode is not a visible error; it is a sequence that is coherent, plausible, and aimed slightly to the side of what you were teaching.

Keep these yourself. Curriculum goals, the intellectual sequence of a unit, source selection and validation, high-stakes grades. Diagnosis of student needs, special-education decisions, discipline, safeguarding, counseling. Judgment on politically or culturally sensitive matters, parent communication that carries conflict or consequence, determinations of cheating, and any decision that materially affects a child’s educational opportunity. Delegating final responsibility here hands over accountability, and the tool cannot hold any.

The third list is the one worth being stubborn about. Nobody minds a machine drafting rubric language; the call on whether a kid gets a second attempt is the job, and it is a good day when the software knows the difference.

For a single assignment the sort gets more specific, and the five tests for any classroom AI use will run that go/no-go once they are up.

A workflow for using AI as a critical collaborator

The AI generates alternatives, questions, and critiques; you hold the goals, the sequence, the factual integrity, and everything you know about the students in front of you. That stance is the critical collaborator, and it lines up with what the teacher–AI co-design literature describes. Five steps put it to work.

Write the design brief before you prompt. Put down the standard, the enduring understanding, the objective, and what you will accept as evidence of mastery. Add what students already know, the misconceptions you expect, the sources you require, the reading level, the time, and the materials. Then state plainly what the AI must not do. A brief this specific is the difference between a lesson shaped by your unit and one shaped by the average of the internet.

Ask for alternatives, then make it argue with itself. Request three lesson openings, two contrasting explanations, competing assessment approaches. Variety is what makes evaluation possible, where a single polished answer invites passive acceptance. Then turn the tool on its own output. What does this assume about prior knowledge, and where would it produce superficial participation? Which parts drift from the stated objective, what claims need verification, and what would work better without technology at all?

Verify. Facts, quotations, citations, standards alignment, reading demands, developmental fit, accessibility, cultural framing, safety, and whether the assessment measures what it claims to. This is the step that gets skipped when a draft arrives looking finished, and looking finished is not evidence of anything.

Localize. Add what the system cannot know: your class history, the relationships in the room, the misconception a few students hit last Tuesday. Add the community context, the materials you actually have, what your school expects, and where you are in the unit. Localization is what turns a correct but generic plan into a lesson.

Pilot one component, then look at what changed. Test a discussion protocol or a question set or a feedback workflow before redesigning a course around it. An EEF/NFER randomized trial with 259 teachers across 68 English secondary schools studied ChatGPT-supported planning in Key Stage 3 science. Teachers with access spent 56.2 minutes a week on the target planning tasks against 81.5 for the control group, a drop of about 31%. A blinded expert panel detected no apparent fall in resource quality. Those teachers used AI for one or two components rather than whole lessons. So pilot narrowly, then ask the two questions the trial could not: did students learn more on their own, and where did the saved time go.

The design brief is the step everyone skips and the one that does the work. Ten minutes writing down what you will accept as evidence of mastery is what keeps the plan pointed at your unit rather than at the average of the internet.

Habits of mind that help, and ones that get in the way

Treating AI as an oracle encourages acceptance of plausible unverified output and quietly moves responsibility for curriculum off the teacher. Treating it as a neutral tool ignores that its answers depend on training data, system design, provider rules, and probabilistic generation, in a way a pencil’s do not. Treating it as primarily a time-saving device makes the saved time the measure, when the measure is whether instructional quality held and where that time went.

Two framings do more damage than any of those. “Students will use it anyway, so every assignment should incorporate it” mistakes inevitability for a pedagogical reason; students also encounter calculators and search engines, and that settles nothing about any particular task. And the framing that the only options are prohibition or unlimited access skips the middle ground entirely: use that is teacher-controlled, role-constrained, and specific to one phase of the work. That middle is worked out at the assignment level in the Four Modes of AI Assignments.

The productive habits are the mirror image. Treat every output as provisional, a draft or a possibility or a hypothesis. Define the cognitive role you want before you ask, so the tool arrives as critic or translator or practice partner instead of as a general-purpose answer machine. Assume a fallible participant, since you remain responsible for the truth and the consequences of the work either way. Fade the scaffold as competence rises, for yourself as much as for students. And treat human relationships as the instructional infrastructure: the reason to hand off a rubric draft is that discussion, observation, mentoring, and feedback are what you would rather be doing with the hour.

How this looks in Kindred K-12

The workflow above is how the assistant in Kindred K-12 is built to be used. You bring the objective and the expertise, ask for options rather than an answer, then verify and adapt what comes back — building the materials for a lesson and adapting the sources you already teach with. What it earns you is not faster materials. It is time back for the parts of teaching that need a person in the room.

Kindred K-12 sits on the teacher’s side of the work: no student accounts, no student personal information, and the assistant never chats with students. What you make with it does reach students, and what it says is your call.

If this is the kind of thing you want to keep thinking about, the Kindred K-12 newsletter carries more of it, and the rest of the collection is at AI in K-12.