Reverse-Question the AI
Everyone teaches you to prompt better. This playground trains the opposite, more valuable habit: interrogate the output before you trust it. You're shown a seeded, anonymized AI claim — a plan, a diff, an analysis — and instead of grading its answer, you write the one reverse-question you'd put back to the model. A deterministic rubric scores your question on four axes — is it specific, verifiable, scoped, and does it challenge the assumptions — then reveals what a strong reverse-question would have caught. Runs entirely in your browser (0 uploads, works offline).
How to use it
- Read the AI output card. Pick a scenario — AI / Eng, Product, or Marketing — to swap in a different confident claim.
- Type the single reverse-question you'd ask the model back before acting on that claim, or load a preset — Vague, Sharper, or Sharp — to see the difference.
- Watch the sharpness score and the four-axis checklist update live; each unmet axis tells you exactly what your question is failing to pin down.
- Open the X-ray panel to compare against a model reverse-question and see the planted flaw it catches. Copy or share your score.
What this clears up (the fundamentals)
- A good answer starts with a good question about the answer. Confident-wrong output is the default failure mode of LLMs; the fix isn't a cleverer prompt, it's a sharper follow-up that forces the model to defend the claim.
- "Sharp" is measurable, not a vibe. It decomposes into concrete properties: names something specific, asks for evidence, stays scoped, and pokes the premise. Score each and vague curiosity becomes an interrogation.
- Verifiable beats plausible. "How do you know?" and "what's the source?" separate a claim you can check from one you're just hoping is true.
- Challenge the assumption, not the wording. The highest-leverage reverse-questions surface the thing the model assumed — the edge case it skipped, the benchmark it borrowed, the premise it never stated.
Where it's used
A FREE, top-of-funnel critical-thinking drill — the fastest way to feel why trained distrust beats prompt-craft. It's a miatz build-lab concept playable: play it here, then learn to build the specificity rubric and the signature reveal yourself. The same rubric grades reverse-question items in the paid trivia engine and in Tuukul's Accept-gate drills.
FAQ
What is a "reverse-question"?
The question you ask the AI about its own answer before you trust it — the follow-up that makes the model justify, cite, or narrow its claim instead of restating it. Reversing the direction of inquiry is the core habit this trains.
How is the sharpness score calculated?
Your question is checked against four boolean axes — specific, verifiable, scoped, challenges-assumptions — using simple keyword and length heuristics. The score is round(axes met / 4 × 100), so 0, 25, 50, 75, or 100. It's a deterministic rubric: no model call, no randomness.
Why did my question score low?
Usually one of three things: it's too broad ("tell me about…"), it asks for a summary instead of evidence, or it accepts the premise instead of probing it. The checklist and X-ray name the exact axis you missed and how to fix it.
Does a high score mean my question is perfect?
No — it means it clears four useful bars. A real reverse-question also depends on domain judgment the rubric can't see. Treat the score as a fast coach for the shape of a good question, not a grade on its content.
Is anything uploaded?
No. Your question is scored entirely in your browser — nothing is transmitted, stored, or logged. Turn off your Wi-Fi and it still works.
Limits
A teaching rubric with four fixed axes and keyword heuristics — it rewards the structure of a sharp reverse-question, not its factual aim. A question can score 100 and still miss the point of a specific claim, or score low while being exactly right for the context. The habit it drills (interrogate before you adopt) is the real practice; the number is just a coach.
Related
Part of the Demystify Playgrounds. Explore the rest from the Playgrounds home.
Bookmark this page (Ctrl+D, or ⌘D on Mac) — it works offline the next time you need it.