The golden rule
A brief is well written when a generic AI answer is insufficient by design. AI is not prohibited — prohibiting it is unenforceable and contradicts the premise. Problems are written that AI alone cannot finish.
And the negative test: if you can name the right answer to your brief, the brief is not finished. A good brief has several valid solution spaces.
The engine: the five movements
A brief that engages does not do it through the topic — it does it because it forces these movements, and each movement is another pass over the learning object:
| Movement | What it forces |
|---|---|
| 1. Create your own version | The participant does not consume the learning object: they produce THEIR version of it — their map, their timeline, their ecosystem, their problem. The own version is what forces understanding. |
See moreThe distinction that holds the whole brief together: consuming the learning object (reading it, summarizing it, reciting it) is not the same as producing your own version of it. By producing, the participant has to decide what matters, what goes in, what is left out and how each part relates — decisions only someone truly looking at the learning object can make. It is Piaget's idea (we construct our version of the world) and Papert's (we learn by making something that can be shown and discussed), turned into a design rule: our position is that if the brief does not force producing a version of one's own, it is still consumption with extra steps. And the Challenge canon says it straight: a brief's statement IS a Learning Object — with identified objectives, not a hackathon prompt. Quality bar: will each team's version necessarily come out different? If they will all look alike, the brief is asking for reproduction, not an own version. | |
| 2. Iterate until it works | The AI’s first result is not enough (if it is, the brief is badly written). Ask again, adjust, direct — each iteration is another pass over the learning object. |
See moreThe AI's first answer is working material, not a result. Each lap — asking, reviewing what came out, adjusting the instruction, asking again — is another pass over the learning object, and that is where the learning accumulates. It is deliberate practice in new clothes: repetition with effort at the edge of one's own capability, with feedback that does not flatter. The strongest field evidence we have (the guardrails experiment, 2025) shows the reverse: when the AI delivers the finished answer, performance rises in practice and learning collapses at the no-AI exam. Iteration protects against exactly that. Quality bar: if the brief can be solved with one well-written request, raise its openness or its friction — it is still an errand. | |
| 3. Verify outside the model | Something in the brief forces leaving the AI: an authorized source, a person, an observation, a measurement. The AI gets things wrong with apparent confidence — discovering that is learning. |
See moreAt some point in the brief, the team has to leave the AI: an authorized source, a real person, an observation, a measurement from the classroom or the neighborhood. That exit is the friction point made activity — the place where the model, predictably, gets it wrong with apparent confidence. Every error detected against a source is domain knowledge gained: whoever discovers the AI put the peak in the wrong range has just learned where the peak goes. That is why local information is asked of the team and never invented in the statement. Quality bar: name the verification source in the brief. If you cannot name what it is verified against, the movement is not designed. | |
| 4. Compare | Two sources, two versions, two contexts, two accounts. Comparison forces looking at the learning object twice and explaining the difference. |
See moreTwo sources that do not agree, two versions of the same work, two contexts, two accounts of the same fact. Comparison forces looking at the learning object twice — and explaining the difference, which is where judgment shows. A real discrepancy is design gold: authorized sources giving different altitudes, the textbook and another source telling the period differently. The team has to decide which to use and defend the why — judgment, not obedience. Quality bar: the comparison must demand a decision (“which do you keep, and why?”), not just a table of similarities and differences. | |
| 5. Review and critique | Detecting what is wrong in what was produced — one’s own or the AI’s — and being able to say why. Critique with judgment is the evidence that learning happened. |
See moreDetecting what is wrong — in what the AI produced and in one's own — and being able to say why. Critique with judgment is the observable evidence that learning happened: nobody can point out a map's error without knowing geography. It is also what the Challenge rubric reads: human-AI direction is judged by decisions (what was rejected, what was verified, what was transformed), never by volume of use. And the no-AI defense is this movement's final test: explaining what was reviewed with the tool switched off. Quality bar: the brief must produce criticizable moments — intermediate versions, plausible errors, recorded decisions. A brief that only generates a polished final product leaves nothing to review. | |
The bar: if your brief can be solved with one request to the AI and zero verifications outside the model, it is still a retrieval task — the same one a search engine has solved for three decades. Level it up with the movements it is missing.
The nine properties — checklist
Every Challenge brief satisfies all nine before publishing. Use them as design questions:
| # | Property | The question you ask yourself while writing |
|---|---|---|
| 1 | Authentic context | Is there a real situation with real people affected — or is it a textbook statement? |
| 2 | Incomplete information | What datum do I withhold on purpose, so the team has to go find it? |
| 3 | Values in tension | Which two good things cannot both be maximized in this problem? |
| 4 | Local or live information | What must the team supply that no model has — an interview, an observation, a classroom fact, a measurement? |
| 5 | Named consequences | Who pays the cost of each possible decision? Can I name them? |
| 6 | Demonstrable artifact | What can be TRIED live — not presented, tried? (Any medium counts; none is mandatory.) |
| 7 | Perturbation | What change of conditions, delivered mid-work, stresses the design without destroying it? |
| 8 | Creation record | Which consequential moments will need documenting — human decision, AI contribution, what was rejected and why? |
| 9 | Defense | What questions would I ask, with no AI at the table, that only whoever really did the work can answer? |
The friction point — the secret field
Every brief declares (in your authoring record, not to the participant) where the AI will fail, predictably: the datum it does not have, the detail it will smooth toward the average, the operation it will do wrong with total confidence. A brief without a declared friction point is not a brief: it is an errand. The friction seeds property 4 — the live information the team will have to supply.
The ethical dilemma goes inside, not glued on
The brief carries an ethical decision the team cannot escape — and it is structural: whoever ignores it must end with a visibly worse solution, not a missing paragraph. Patterns that work: the easy technical fix is the privacy-problematic one · the most personalized help demands the most intimate data · accessibility raises cost, and cost excludes someone · the persuasive campaign risks caricaturing the community it serves.
Anti-patterns — what is NOT a brief
The author’s failure is worse than the team’s: do not write an “AI task with a stopwatch”. Rejected forms, verbatim from the canon:
- “Ask the AI for a business plan.”
- “Create the best generated image about climate change.”
- “Use a chatbot to write an essay.”
- “Make a website about healthy eating.”
- “Create the best prompt.”
Also: no medium is mandatory (requiring working code is an equity failure) · no ceremonial templates teams must “perform” · nothing that rewards prompt counts, tool counts or the newest model · never identifiable personal data in AI systems — scenarios are built with public, fictional or authorized information.
Worksheet — write your brief
The format of a rehearsal brief, to fill in by hand or as a team. (The guiding-team section is not shown to participants.)
Brief title
Subject / grade · Real learning object (what you want them to learn — not the artifact)
Learning / curricular objectives (what the participant will know and be able to do at the end; anchored to the curriculum where it applies — if the brief can be solved without touching them, the brief is badly designed)
Authentic context (the real situation, the people affected)
The brief (what own version of the learning object they will create with AI — and what makes it impossible to solve with a generic answer)
Local or live information the team must supply (property 4)
The tension they cannot dodge (the two goods in conflict, and who pays the cost)
Possible output formats (2–4 options; the team picks ONE and declares it at the Divergence Gate — none is mandatory, declaring one is)
What the deliverable must contain, whatever the format (3–5 concrete elements)
What is demonstrated at the defense (demonstrable live, in any medium)
Which knowledge stages the brief forces, and where (research · source validation · data preparation · insight · insight validation · presentation · AI audit)
The rounds the brief forces (what does it demand anticipating? against what is it audited? what makes redoing inevitable?)
— Guiding team only — Friction point: where will the AI fail, predictably?
— Guiding team only — Perturbation (mid-work; stresses without destroying; the same for everyone)
Quality test in a handful of questions: can the AI alone finish it? (if yes, friction is missing) · can I name THE answer? (if yes, openness is missing) · how many of the five movements does it force? (fewer than three: level it up) · does the ethical dilemma change the design or is it decoration? · will the team know, passing the Gate, WHAT will exist at the end — not just the concept, the format? · can the brief be solved WITHOUT touching the learning objective? (if yes, redesign it) · does the brief force crossing the knowledge stages — research, source validation, data preparation, insight generation and validation, presentation, AI audit — in a way that can be pointed out? · does it leave room for the team to imprint its own relevance — its classroom, its neighborhood, its experience?
Framework derived from the aiLearning Challenge brief-authoring specification (Smoother canon). Educators’ version prepared for the rehearsal.