How to use AI in lesson planning effectively

The good news about using AI in lesson planning is that it can genuinely save you time. The less good news is that the time savings only show up when you stop treating AI as a magic resource generator and start treating it as a slightly unreliable colleague who is good at first drafts but needs supervision.

The workflow that tends to work is the same three-step loop: Generate, critique, adapt. You ask the model for a draft, you read it sceptically against what your class actually needs, and you adjust both the resource and the prompt before using either in a classroom. The teachers we have spoken to who report the biggest gains run this loop fast and refuse to skip the middle step.

This guide walks through the workflow in practical detail, with the prompts and patterns that hold up under classroom pressure. It is grounded in current thinking about AI in education, including the pedagogical principles that underpin Oak's Aila assistant. AI tools are moving quickly enough that some of the specific advice here may shift inside a year. The underlying approach should hold longer.

Before getting into workflow, it helps to be honest about which planning tasks AI tends to help with and which it does not.

Good candidates for AI assistance include drafting retrieval starters from a list of topics, producing differentiated versions of a worksheet you have written, generating multiple worked examples of the same technique, writing knowledge organisers from your key points, and turning a rough lesson outline into a structured plan. AI is doing something near the centre of its competence: Producing variations on a clear input.

Weaker candidates include deciding what to teach, sequencing content across a half-term, designing a unit that responds to your specific class, and producing accurate exam-style questions for high-stakes use without human review. These require expert judgement, accurate spec knowledge, or both.

The practical implication is that the most useful AI workflows save time on execution rather than thinking. You decide what the lesson needs to do, AI helps you produce the materials faster, you check and adjust.

The fastest way to get poor AI output is to ask it to plan a lesson from a single sentence. "Plan a Year 9 lesson on photosynthesis" gets you a generic, off-spec plan that will need a lot of editing.

Do your planning thinking first, then use AI to produce the artefacts. The thinking that needs to happen first: One concrete learning outcome, the prior knowledge you can assume, the most likely misconception, the kind of practice you want, and the assessment for learning move you will use at the end.

That is the same thinking you should be doing anyway. AI does not remove the need to plan. It removes the need to draft. Teachers who try to use AI for the thinking end up with lessons that are pedagogically average and curriculum-mismatched.

Most teachers find building a small set of reusable prompt templates pays back faster than almost anything else. A full lesson prompt template includes: The class context (year group, exam board, ability range), the topic and learning outcome, the prior knowledge you are assuming, the likely misconception, the pedagogical structure you want (retrieval starter, modelling, guided practice, exit ticket), any representation or tone requirements, and the format you want the output in.

When written out, this looks like a long prompt. That is the point. The trade-off: An extra two minutes writing the prompt for ten minutes saved editing the output. Once you have a template, filling in the variables for a new lesson takes maybe three minutes.

Here is the kind of prompt structure that tends to produce material you can actually use, with a worked example for a sociology lesson.

Class: Year 13 sociology, AQA spec, mixed prior attainment.

Lesson topic and outcome: Functionalist view of education. By the end of the lesson, students should be able to identify the key claims of Durkheim and Parsons on education and evaluate them with reference to one criticism each.

Prior knowledge: Students have covered the basics of functionalism in Year 12 and introductory work on the role of education last week.

Likely misconception: Students often confuse functionalist and Marxist views on education and conflate "meritocracy" with "equal opportunity".

Lesson structure I want: Five-minute retrieval starter, ten-minute teacher explanation of Durkheim and Parsons, a worked example of an evaluative paragraph, fifteen minutes of guided practice in pairs, and a five-minute exit ticket.

Other requirements: Varied range of student names. Reference one contemporary critique alongside the classical ones. Pitch the language at the level of a strong A-level cohort.

Format: Produce the lesson plan as a table with timings, teacher actions, and student actions. Then produce the retrieval starter, the worked example, and the exit ticket as separate outputs.

Whatever the model produces, treat it as a draft. A fast critique turns a 70% good draft into a 95% good lesson without much work. A slow critique can take longer than writing the lesson yourself would have done.

A workable critique runs through three filters. First, accuracy: Is everything true and aligned to the specification? Second, level: Is the pitch right for this class? Third, pedagogy: Are the moves the lesson uses the moves you would have chosen, or has the model defaulted to something generic?

Most AI-generated lesson plans fail on the third filter more often than the first two. Models tend to produce structurally fine but pedagogically average lessons. They underuse retrieval, rarely include hinge questions, and tend not to plan for misconceptions. The critique is where you put those back in.

This step is the one most teachers skip and the one with the biggest pay-off over time. When you find yourself making the same edit across multiple lessons, the prompt template needs updating.

Keep a short note of recurring edits. After a fortnight or so, update your template to bake in the fixes. Common ones include: Adding an instruction to include a hinge question, asking for differentiated versions of any practice task, specifying exam board terminology, and requiring varied names and contexts.

A template that has been through two or three iterations tends to produce output that needs much less editing. Teachers who feel AI is not saving them time are usually using the same generic prompt they used in week one.

Some lesson planning tasks have a particularly good ratio of time saved to risk introduced. These are good candidates for early AI workflow investment.

Differentiation. Producing scaffolded versions of a worksheet for students who need more support, and stretch versions for students who finish early, is one of the most time-consuming parts of preparation. AI tends to be good at this when given a clear baseline and a description of the adjustments needed.

Retrieval starters. Asking the model to produce five quick-fire retrieval questions from a list of topics you covered earlier in the term is fast and reliable. Verify the questions, but the production cost is lower than writing them yourself.

Worked examples. For maths, sciences, and structured-essay subjects, AI can produce additional worked examples to extend the range students see. Always work through the maths yourself.

Scaffolds and knowledge organisers. Sentence stems, paragraph frames, structured tables, single-page summaries from a set of key facts: AI is good at producing variations on these from a clear example. Layout still needs your eye.

Equally important is being honest about where AI does not yet help much, or where the risks outweigh the benefits.

Unit planning and sequencing. Deciding what to teach when, how it builds on previous learning, and how it sets up future learning depends on knowing your students, your department, and your specification deeply. AI can produce a generic scheme of work, but the sequencing decisions tend to be the bits it gets least right.

Assessment design. AI-generated exam questions look plausible and are often subtly off-spec. For high-stakes mock exams or end-of-unit assessments, treat AI as a starting point at best and check every question against your board's mark schemes.

Responding to a specific class. Which student needs a different question, which pair to put together, which misconception is worth a whole-class detour: AI cannot help here because it does not know your class.

Feedback on student work. Generic AI feedback can work as a starting point, but the act of reading student work and noticing what the student actually needs is where the teaching judgement happens. Outsourcing that flattens the responsiveness that makes feedback effective.

If you read the public material on Oak's Aila, one thing that stands out is how much pedagogical structure is built into the underlying prompts. Aila enforces certain moves: Retrieval, clear learning outcomes, modelling, varied practice, formative assessment. The structure is doing a lot of the heavy lifting.

You can replicate this in your own prompts. Worth defaulting to: Explicit retrieval practice, worked examples with reasoning made visible, hinge questions that diagnose specific misconceptions, ramped practice, and exit tickets that retrieve the key idea. All of them are routinely missed by generic AI output unless you ask for them. Building these into your default template is a one-time investment that pays off every time you use it.

For science and maths in particular, the strongest workflow for many teachers combines AI-generated material with a curated bank of vetted content. Cognito's exam-aligned quiz library and topic videos give a reliable baseline that AI can build around, rather than asking AI to produce high-stakes content from scratch. The point is not that one approach replaces the other. It is that mixing them lets you spend AI time where it adds most value (variation, scaffolding, retrieval) and lean on vetted material where accuracy matters most.