Thirty Minutes of Planning and Where AI Actually Belongs

It is 4:45 on a Wednesday and three people are left in the staff room. Ms Chan, who teaches P5 General Studies, opens next week's scheme of work and says to the colleague at the next desk: "Community planning, period five next Wednesday, and I still need three worksheets." What she has in front of her is one page of unit learning points, a textbook, and a list of twenty-seven names.

Planning is slow because of the number of decisions, not the amount of typing. AI can shrink the typing to almost nothing, and the thirty minutes that are actually worth something should go entirely to the judgements only you can make.

Where the thirty minutes go

Colleagues who conclude that AI planning "doesn't work" have usually spent all thirty minutes arguing with the model, asking it to be "more specific" over and over, and ending up with a plan that contains everything and fits nobody.

The first eight minutes generate nothing

These eight minutes are for writing down what the model cannot possibly know. Copy the learning points verbatim from the textbook. Describe the class honestly — twenty-seven, upper-middle, five weak in computation, two non-Chinese-speaking, one pupil with ADHD in the front row. Say where the last lesson landed and where it stuck. Ms Chan's previous lesson covered classifying community facilities, and she noticed that "planning" makes her pupils think only about putting up buildings. That one sentence has to go in, or the starter the model writes will begin from zero and burn the first ten minutes of the period.

Fourteen minutes, four separate asks

Four asks: plan skeleton, slide sequence, three tiers of worksheets, follow-up questions. One thing each time, each building on the last. Asking for plan-plus-worksheets-plus-slides in a single request reliably produces three half-finished artefacts.

The last eight minutes are for cutting

Whatever comes back will be too long and the activities will overrun. This is the cutting pass: five activities down to three, "ten-minute group discussion" down to "two minutes with the person next to you", and every off-curriculum term corrected. Cutting is far faster than writing, and that asymmetry is where the time saving actually comes from.

The context paragraph

The context worth eight minutes is not a one-line introduction saying you teach P5 General Studies. It is four concrete things: the unit's learning points as written, the class as it really is, where the previous lesson ended, and the number of minutes you genuinely have. A thirty-five-minute period, after collecting books and settling a class that has just walked in, is usually twenty-eight.

Write those four once and they carry the whole unit. On a school platform they already exist as fields — subject, level, unit, objectives — so the teacher selects rather than types. On a general chat tool, keep the paragraph in a file and paste it each time.

One ask at a time

Skeleton before prose

The first ask is the skeleton only: segment names, minutes, which learning point each one serves, what the teacher does, what the students do. Explicitly forbid scripted narration and lead-in lines. A skeleton is a dozen lines, you read it in thirty seconds, and you catch a broken rhythm immediately instead of discovering on page two that the starter has eaten half the lesson.

For the S2 Geography lesson on urban renewal, the first skeleton gave eighteen minutes to a Kwun Tong case study. One glance was enough — cut to eight minutes and two street-level photographs, before and after, then move on.

Answers arrive with the questions

Flesh comes second. Worksheets must be generated together with their answer key and the learning point each item serves, or you end up working through the whole sheet yourself just to find out whether it holds up. Ask for answers aligned to item numbers, then spot-check three of them. Calculation and data-reading items are where stale keys hide.

Real tiers and fake ones

The classic fake tier is five questions, then eight, then twelve. Quantity is not difficulty. A struggling pupil who does five questions is still stuck; they have simply copied less.

Real tiers differ in what the student has to think. On community planning, the foundation tier gets a neighbourhood map and labels three facilities with one use each. The core tier gets the same map plus a short resident complaint and identifies what is wrong with how the facilities are spread. The challenge tier proposes a use for a vacant lot and names one stakeholder who would object, with a reason. Same material throughout, with scaffolding stripped away layer by layer — a glossary and sentence openers at the foundation, glossary only at the core, nothing at the top.

Say this outright when you ask: all three tiers use the same material, the difficulty comes from the thinking move, and adding or removing questions does not count.

Whether your tiers are real is not measured in question counts — it is whether the weakest tier finished having thought about something once.

What the panel keeps

When one colleague tunes a context paragraph and a set of generation rules that work, leaving it on a personal laptop means the other seven people in the panel each rediscover it alone. Saved as a panel template, the difference shows next term: a newly joined teacher's first plan already has plausible timings and the right vocabulary.

A panel head reviewing plans is not counting words either. Three things matter — no learning point quietly dropped, timings that match a real period, and tiers that genuinely tier. Under a shared template, that review takes a few minutes.

Common misreadings

• The claim: Planning with AI outsources the plan, and the teacher just presses a button. What actually happens: Every judgement stays with the teacher — which activity to cut, how far to push the questioning, which pupil needs which support. The model writes up decisions you have already made.

• The claim: Longer, more detailed prompts produce better output. What actually happens: Overloaded prompts make the model trade one requirement against another and do all of them thinly. Four separate asks, one artefact each, is far more stable.

• The claim: Thirty minutes applies from the first attempt. What actually happens: The first run takes about ninety minutes, because the context is being built from nothing and you are still learning which instructions need to be nailed down. Thirty minutes becomes true after three or four units.

Three things to try on Monday

1. Pick one lesson from next week and spend eight minutes writing four things into a file: the learning points as printed, the class as it is, where the last lesson stopped, and the real number of usable minutes. Reuse that file for the whole unit.
2. Generate the skeleton only — no worksheets, no narration. Add up the minutes it proposes, compare against the minutes you actually have, and cut segments until the two match.
3. Take a worksheet you already use and ask whether the weakest tier thought about anything. If not, rewrite one item: same material, different thinking move.