Teaching Principles

transfers unchanged

1. Write the questions the course answers, before writing anything else

Boxer's first move is to fix a small set of core questions a unit must leave a learner able to answer, made as specific as the subject allows. Everything after that is judged against them.

The failure this prevents is the ordinary one: building a course out of the materials that happen to exist. Sequence from the content, never from the resources you have. A slide deck that exists is not a reason to teach what is on it.

In our material: every 101 deck opens with what the session answers, and the deck standard we check in CI exists to stop a deck drifting into a tour of a topic.

transfers unchanged

2. Separate knowing the subject from knowing how to teach it

Subject knowledge and pedagogical content knowledge are different things, and having a great deal of the first is no guarantee of the second. Knowing what a Gini coefficient is does not tell you which order to introduce it in, what people reliably get wrong about it, or which worked example makes it land.

In development and policy work this cuts particularly hard, because most people teaching the subject arrived from practice rather than from teaching, and practice builds the first kind of knowledge almost exclusively.

adapted for social science

3. Identify prerequisite knowledge, then check it rather than assume it

Before teaching anything, name what a learner must already hold to make sense of it, write questions that detect whether they hold it, and ask those questions at the start. Learners are not blank slates; the risk is not that they know nothing but that what they know is partial, wrong, or attached to the wrong thing.

What changes here. Science prerequisites are largely hierarchical — you cannot do moles before you have particles. Social science is much flatter: you can teach caste before or after stratification theory and both orders work. So the prerequisite question shifts from "have they met this yet" to two other things. First, quantitative floor: percentages, rates versus counts, real versus nominal, what a median is and why it differs from a mean. This is the prerequisite most consistently assumed and least often checked, and its absence is invisible until a learner reads a chart backwards. Second, institutional furniture: what a Gram Panchayat is, what a line department does, what a scheme guideline is. Learners from outside the country, and often from inside it, do not have this and will not say so.

In our material: our courses assume no economics or statistics background, and the reading companions carry the quantitative floor explicitly rather than leaving it to be picked up.

transfers unchanged

4. Decide, deliberately, how people are meant to respond

Boxer calls this the means of participation, and his point is that it is almost always left implicit: where does the answer go, what if you do not know, does it need a full sentence, are you writing it or saying it? Learners who cannot tell will default to not participating, and the teacher will read silence as comprehension.

Related, and worth holding separately: participation ratio is what proportion of the room is doing something; thinking ratio is what proportion is thinking. They come apart constantly. A lively discussion among four people has a high thinking ratio for four and near zero for everyone else. Copying notes has a high participation ratio and almost no thinking ratio at all.

In our material: this is the hardest principle to honour in self-paced online learning, where there is no room to read. It is why our courses carry reflection prompts and worked examples rather than only slides, and why the studios ask for a produced artefact rather than a completion click.

transfers unchanged

5. Novices are not small experts, and you cannot feel your own expertise

An expert's knowledge is organised differently from a novice's, not merely larger. The consequence, which Boxer calls the curse of knowledge and the research literature calls expert blindness, is that the steps you skip are invisible to you precisely because you have automated them. You cannot introspect your way out of this; you find the missing step by watching someone fail at it.

This is why we test explanations on people who do not already know the answer, and why a subject expert reviewing their own material is the weakest form of review available.

transfers unchanged

6. Teach what a thing is and what it is not

A concept taught only through examples that fit it leaves its boundary undefined, and learners will draw the boundary somewhere arbitrary. Non-examples are how you elicit boundary conditions, and they must be chosen rather than improvised: the useful non-example is the one that is close enough to be tempting.

Social science needs this more than science does, because its central terms are elastic and have been stretched by usage. "Empowerment", "participation", "sustainability", "capacity building" and "community" are all words a learner can use fluently for a year without holding any boundary at all. Teaching participation without a worked non-example — a consultation whose outcome was fixed in advance — leaves the word decorative.

In our material: the course lexicons exist for this. A definition alone is the weakest form of the entry; the useful part is the pair of cases either side of the line.

transfers unchanged

7. Build the diagram in front of them; never read your own slide aloud

Words and pictures together beat either alone, which is the multimedia effect. Two failure modes sit either side of it. Split attention is when related information is separated so the learner has to hold one part while hunting for the other — a diagram here, its key there. Redundancy is when the same information arrives through two channels at once, which is exactly what happens when a presenter reads text off a slide while the audience reads it too.

Both are solved by the same move: start from a blank canvas and construct the diagram while talking, so each part arrives once, in the order you are speaking about it. It is worth rehearsing.

In our material: this is the reason our decks carry built-up diagrams rather than a finished graphic with a paragraph beside it, and the reason slides carry cues rather than the script.

transfers unchanged

8. Directions of travel

Boxer's most immediately usable chapter is a list of routes through an explanation. We use all six, and naming them is what makes them choosable rather than accidental:

• Wood → trees → wood. Show the whole, go into the part, come back out. The return trip is the one people skip, and it is where the point lives.
• Simple → complex. Teach a simplified version knowing it is incomplete, and say that it is. Some concepts take years.
• Example → principle → example. Usually faster than starting from the principle, because the first example gives the principle something to attach to.
• Explanation → definition. Definitions are compressed and therefore hostile to a novice. Explain the thing, then define it, then have them use the definition.
• Hinterland → core. The story around a piece of knowledge — who found it out, what they were arguing with — can set the stage. Used carelessly it becomes the lesson.
• Conflict → resolution. Set the content up as a problem with stakes. Stories are cognitively privileged and this is the cheapest way to use that.

Explanation → definition and example → principle → example do the heaviest lifting in our material, because the terminology is where social science loses people.

transfers unchanged

9. Analogies pass a checklist or they do not get used

An analogy has a source the learner knows, a target they do not, and an inference connecting them. Boxer's conditions are worth applying literally. Use one only if the learner really is familiar with the source; the leap is believable; it is not distracting; it does not seed a misconception; and it is actually necessary. A decorative analogy still costs working memory.

The failure specific to our subject is the analogy that smuggles in a politics — comparing a national budget to a household budget, for instance, quietly imports the assumption that a government cannot run a deficit, which is the conclusion under discussion rather than a neutral aid to it.

where we part company

10. Misconceptions in social science are positions, not errors

Boxer's account of misconceptions is that they are plausible and predictable, that you should anticipate them from your own knowledge of the domain, and that they are hard to shift precisely because they are reasonable. He is candid that research offers little practical advice beyond knowing what they are and reducing abstraction.

All of that holds for us. What does not hold is the underlying model. "Respiration means breathing" is a mistake a learner will drop once shown; it costs them nothing to give up. "People are poor because they do not work hard" is not that. It is attached to a person's account of their own position, of what they deserve, and of what they owe. Treating it as a factual error to be corrected produces public agreement and private retention, which is the worst outcome available — it removes the disagreement from the room while leaving it intact.

What we do instead: give the position its strongest form before touching it, put the evidence where the learner can work it themselves rather than receiving a conclusion, and separate the empirical claim from the moral one so the two can be argued separately. Where the disagreement is genuinely about values rather than facts, say so and stop. A course that pretends a value disagreement is a knowledge deficit is doing something other than teaching.

In our material: this is why so much of what we build is a dataset or a simulation rather than an argument. A learner who moves the transfer slider themselves and watches what happens has done something an assertion cannot do for them.

adapted for social science

11. Check every word, including the ones that look ordinary

Boxer distinguishes Tier 3 vocabulary — the obvious technical terms — from Tier 2, the general academic words that get assumed. Tier 2 is where the damage is, because nobody thinks to teach it.

Our subject has a third category that science does not: words that carry an everyday meaning and a technical one that differs from it. Discrimination, capital, rent, elasticity, significant, bias, formal, informal, marginal. A learner who maps "significant" onto "important", or "rent" onto what they pay a landlord, will follow an entire argument and take away the wrong thing without ever being confused enough to ask. These have to be taught as false friends, explicitly.

Teaching vocabulary also adds to task quantity, so it is not free. Decide which words matter and let the rest arrive later.

transfers unchanged

12. Understanding is connection, and it is a spectrum rather than a state

Boxer's working definition, drawing on Bartlett's schema theory, is that understanding is new knowledge made solid plus the extent of its connections to what a learner already holds. It is not a switch that flips. It also is not the feeling of understanding, which is why self-report is close to worthless as evidence: an idea can feel like it makes sense and be neither solid nor connected to anything.

The consequence for course design is that we are in the business of building structures over months, not delivering items. A single well-sequenced session is worth more than a complete syllabus badly ordered.

transfers unchanged

13. Check, then consolidate, then practise — in that order

New knowledge immediately after an explanation is hazy and will not survive being used. Checking for understanding and consolidating it are different jobs: the first tells you whether it landed, the second makes it solid, usually by making the learner say or write it in their own words. Both come before independent practice, or the practice rehearses the wrong thing.

The quick version Boxer describes is worth stealing wholesale: take a diagram the learner has seen, strip the labels, replace them with letters, and ask what each is, what it does, how it differs from its neighbour, and what would happen without it.

transfers unchanged

14. Retrieval is a routine, not an activity

Bringing knowledge back out of memory strengthens it more than reviewing it does, and the effect is larger when the attempts are spaced out over weeks and months. This only works as a habit: a quiz at a fixed point, questions during explanations, questions inside practice, and homework designed as retrieval rather than as extension.

Two details that matter more than they look. Tell learners why you are doing it, or it reads as testing rather than learning. And normalise being wrong by explaining how and why someone might reasonably have got there — a room where wrongness is costly is a room where you get no data.

In our material: practice packs and the challenge sets are built as spaced retrieval rather than as end-of-course assessment, which is why they revisit earlier material rather than only the most recent.

adapted for social science

15. Distinguish single-answer questions from many-answer ones, and mark them differently

Boxer separates questions with one correct answer from questions with several, and notes that the second kind needs a different way of going over it — sample the answers, say what is right about each and why, put the good version on the board.

What changes here. In science most questions are the first kind and a few are the second. In social science it is close to inverted, and a third category appears that Boxer does not need: questions where the answer is contested and the assessable thing is the reasoning. "Did the mid-day meal scheme work?" has no answer to mark. "What would count as evidence that it worked, and what would that evidence miss?" does.

So our default is to assess the warrant rather than the conclusion, and to be explicit with learners about which of the three kinds of question they are answering. Being unclear about that is how a learner comes to believe that a well-argued position is being marked down for disagreeing.

transfers unchanged

16. Design the practice sequence with the equation, not by feel

"Easy to hard" is not a design. Every question in a set should produce learning, and the way to raise challenge deliberately is to move one term at a time: add steps or unit conversions or a figure they must go and find; withdraw a cue; increase abstraction. Awkward numbers raise abstraction more than people expect.

Question sets often need to be longer than feels comfortable. A learner who has done four of something has not practised it.

transfers unchanged

17. Interleave

Mixing questions from different topics beats doing them in blocks. It breaks the algorithmic routine that sets in when every question on a page has the same shape, it spaces retrieval as a side effect, and it forces the comparison between neighbouring concepts that blocked practice never asks for. It also feels worse to the learner and produces lower scores in the moment, which is why it has to be a deliberate choice made by whoever designs the set.

For declarative material, which is most of ours, Boxer's variation techniques are the ones we use: contrast two concepts that are easy to confuse; ask what would follow if some feature were absent; present a wrong answer and ask why someone would give it; and finish sentences of the form because / but / so from a single stem.

transfers unchanged

18. What helps a beginner hinders an expert

Worked examples, heavy scaffolding and step-by-step instruction help someone new and actively get in the way of someone who is not — the expertise reversal effect. Support has to be withdrawn as knowledge grows, and task quantity raised to fill the gap. Left in place it becomes frustrating and, worse, it makes an experienced learner conclude the material is not for them.

In our material: this is the tension between the 101 courses and the flagship ones. It is also why the studios are goal-free — a brief rather than a procedure — since open tasks of that kind suit people with knowledge and defeat people without it.

transfers unchanged

19. A test samples; write it so the sample supports the inference

You cannot test everything taught, so any assessment is a sample from which you infer something larger. Boxer's distinction is between a literal test, which checks the specific things asked, and a representational one, which is built with enough coverage that performance licenses a claim about the whole. Early on, literal is appropriate. Later it is not, and using it anyway tells you a learner has the four things you happened to ask about.

Hold on to the other half of this as well: performance is not learning. Doing something correctly today, immediately after being shown, is weak evidence that it will be there in a month. Delayed and spaced checks are the ones that carry information.

transfers unchanged

20. Feedback to the class, and analysis by question, beat written comments

Comment-based marking is expensive and, without protected time for a learner to act on it, of doubtful value. Whole-class feedback — noticing an error while circulating, sampling a handful of submissions, or a short check followed by going over it — costs a fraction and reaches everybody.

After any substantial assessment, do the analysis question by question to find the pattern rather than going through the paper item by item. The output is a list: what to reteach, what to put into retrieval, what to do differently next time. Feedback also moves along a spectrum — very directive and immediate at the start, less directive and more learner-focused as knowledge grows.

In our material: the gradebook is built for this. It takes what a studio already produces and makes it markable in a batch, so the instructor's time goes into the pattern across submissions rather than into writing the same comment thirty times.

adapted for social science

21. Ask what the activity is for before building it

Boxer's question about a demonstration is simply: why am I doing this? He allows two good answers. One is awe and wonder, which works because brains are prediction machines and being wrong is memorable — it sets up a conflict → resolution route through the content. The other is to teach specific content, by making something abstract concrete, narrated, supported by a diagram alongside, and followed immediately by questions.

Both answers apply to a simulation. The counterpart of the surprising demonstration is the model that produces a result the learner did not expect from rules they agreed were reasonable — segregation emerging from mild preferences, a commons collapsing without anyone behaving badly, a targeting rule excluding the people it was written for.

The honest note: an interactive thing that has no answer to "why am I doing this" is entertainment, and we have built some. The test we apply now is whether a learner could state what the activity taught them without repeating what they did in it.

transfers unchanged

22. Slow the practical down

Boxer's diagnosis of why practicals so often teach nothing is the equation again: task quantity is high, abstraction is high, the learner's knowledge is low, and support is hard to give while twenty people are doing different things. A worksheet with fourteen steps is not scaffolding, it is fourteen more things to hold.

His fix is the slow practical — break it into small steps and alternate the learner's turn with the facilitator's, so everything stays manageable, problems are visible while they are still small, and the explanation continues throughout. This is wood → trees → wood applied to an activity, and it transfers directly to a data exercise or a field visit. Most of a fieldwork day is spent on logistics; the learning happens in the twenty minutes where somebody says what to notice.

In our material: our labs are built in steps with the reasoning between them, rather than as a single open sandbox with a brief at the top.

transfers unchanged

23. Engagement is a poor proxy for learning

This is the principle most inconvenient to a platform like ours, which is why it is here. A room can be busy, cheerful and occupied and learn nothing, and activity is far easier to observe than learning, so it gets used as the measure. Boxer's warning about gaming for the sake of engagement applies to us more directly than to most people who will read it.

What actually motivates, on his account and on Deci and Ryan's, is success and the sense of getting somewhere: explain why the learning is worth doing — including the argument that the more you know, the easier further learning becomes — and then engineer early success, which breeds the confidence that sustains the harder middle. Novelty does the opposite over any length of time.

In our material: a game earns its place if it teaches something that is hard to teach any other way. Every one of ours is meant to meet that test, and where one does not, that is a defect to fix rather than a feature to keep.

where we part company

24. Values are content here, not noise

The deepest difference between the two subjects is not about method. School science teaches settled content; where it is unsettled, that is a fact about the frontier rather than about the classroom. Much of what we teach is genuinely contested by informed people who share the evidence, and the disagreement runs on values rather than facts.

Teaching it as though a correct answer exists is a category error in two directions. It misrepresents the field, and it teaches learners that the way to handle a value disagreement is to find out which side the authority is on. So a further principle sits over the top of the rest: name the disagreement, teach the strongest version of each position, and be explicit about which parts of it are empirical and which are not. The empirical parts get taught like anything else on this page. The rest gets argued, and the learner is entitled to leave disagreeing with us.

This is not a licence for false balance. A position that misstates the evidence is wrong and gets taught as wrong. The distinction is between a claim about the world and a claim about what matters, and keeping the two separate is most of the work.