SAT Question Studio
Nine SAT-style question types, math and reading & writing, each able to generate endless new variants in a fraction of a millisecond. A lightweight open model, Qwen, explains any one on request.
WHAT QWEN WAS GIVEN
SHOW THE TEMPLATE
What you’re looking at
Every question type starts as a fixed base example, the kind of item you’d see on the digital SAT. Press New variant and the studio swaps the parts that can change: the numbers, the names, the item on sale, the town in the passage. It then recomputes the correct answer and every wrong answer. The highlighted words show exactly what changed.
Generate 12 at once makes the point: a dozen new questions, each with a correct answer key, in about a millisecond, in your browser, with no AI involved.
The trick is the wrong answers
Anyone can change the numbers in a word problem. What makes an SAT question good is the distractors. Each wrong answer is the result of a specific, common mistake:
- In the gym problem, one wrong answer is what you get if you ignore the sign-up fee, and another is what you get if you add it instead of subtracting.
- In the discount problem, one comes from treating “−25% then +8% tax” as a single −17% change.
- In the grammar questions, the wrong verbs agree with the nearby plural noun instead of the real subject.
Each template encodes those mistakes as formulas, so every variant keeps its traps. Open Show the template to see the skeleton a type is built from.
The other trap: variants that don’t make sense
The first version swapped every slot independently, and it showed. It gave you a software designer who “checks every measurement three times,” a wheelchair ramp steep enough to ski down, and a $200 desk lamp next to a $20 graphing calculator. Every answer key was still correct; the questions just weren’t believable, and a real test item has to be.
The fix was to make each context a bundle: a profession travels with evidence that fits it, a ladder comes with ladder-shaped triangles, a rice cooker with rice-cooker prices. Only the truly independent parts, like names and a few numbers, vary freely. Templating isn’t hard because of the math. It’s hard because every combination has to make sense to a sixteen-year-old reading carefully.
Where the model comes in
Press Explain with Qwen after answering, and Qwen3-30B-A3B writes a tutor-style explanation. It’s an open-weight model from Alibaba, running on Cloudflare’s network. It has 30 billion parameters, but only about 3 billion are active for each word it writes, which makes it fast and very cheap: each explanation costs a tiny fraction of a cent.
The important design choice: Qwen doesn’t solve the problem. It’s handed the verified answer, the worked solution, and the mistake behind each wrong choice, all computed by the template, and asked only to explain. Open What Qwen was given to see exactly that. In early testing without those facts, it confidently explained why a student might pick one wrong answer, using arithmetic that didn’t add up. Small models are good writers and unreliable solvers, so the solving belongs in code.
What this means for test prep
Generating unlimited practice questions is now close to free. The hard, human part is still the part templates capture: knowing which mistakes students actually make, and writing distractors that catch them. That’s where teaching experience matters.
Guardrails
Question generation happens entirely in your browser. Explanations go through a Cloudflare human check, with 30 per visitor per day. The server rebuilds each question from its template before asking Qwen, so it only ever explains the studio’s own questions.