Questioncraft: The Art of the Sharp Question
Question Craft
Three real interrogations. In each, the QUESTION did the work — for better or worse. Read what was asked, and what it produced.
Frost Asks Nixon One Specific Question — And Gets "When the President Does It, That Means It Is Not Illegal"
After days of evasive answers, David Frost stopped asking broad questions about Watergate and asked a narrow, specific one about whether a president could authorise illegal acts in the national interest. Cornered into a precise claim, Nixon gave the line that defined the interviews. The broad questions had let him drift; the sharp one left no room to move.
"A vague question lets a witness wander. A specific one nails their feet to the floor."
Source ↗Loftus & Palmer: Changing One Word in a Question Changed What Witnesses "Saw"
In a classic experiment, witnesses watched the same car crash, then were asked how fast the cars were going when they "smashed" / "hit" / "contacted" each other. The verb alone shifted their speed estimates — and a week later, those asked about cars that "smashed" were far more likely to falsely remember broken glass. The question didn't just retrieve the memory; it rewrote it.
"A loaded question doesn't read the answer. It plants it."
Source ↗"Tell Me About Dogs" vs "Compare a Border Collie and a Bulldog for a Small Flat — 3 Bullets": Same Model, Two Worlds
Ask a language model a vague question and you get a vague, hedge-everything answer — true of nothing, useful for nothing. Name the comparison, the constraint, and the format you want, and the same model returns something you can actually act on. The model didn't get smarter. The question did.
"Garbage question in, plausible nothing out."
Which of these questions got the truth, which buried it, and which got nothing? What did the good question DO that the others did not?
Reading
The Question Is the Tool. Shape It Before You Swing.
A question is not a neutral request for information. It is a tool you press against a source, and its shape decides what comes back. Three tests tell you, before you ask, whether a question is sharp.
Test 1 — Open or closed? A closed question can be answered "yes" or "no" and then it's over. "Did Watergate involve you?" — "No." Dead end. An open question forces the source to produce something: what, how, why, in what way. Closed questions are for confirming a fact you already half-know; open questions are for discovering one you don't.
Test 2 — Specific or vague? "Tell me about the crash" invites a shrug. "How fast were the cars going when they hit?" demands a number. Name the exact thing — the moment, the metric, the format — and you constrain the source into a real answer instead of a safe, useless generalisation. Frost's broad questions let Nixon drift; his specific one trapped him.
Test 3 — Clean or loaded? A loaded question smuggles an assumption inside it. Loftus changed one verb — "smashed" instead of "hit" — and witnesses remembered faster cars and broken glass that were never there. The question didn't retrieve the memory; it contaminated it. If your question contains the answer you expect, you'll get it back — whether it's true or not.
The sharpest questions are open, specific, and clean all at once. Run the three tests on your own question before you ask it — especially when the source is an AI that will happily fill any vague or loaded prompt with confident, plausible noise.
- 1
Why did Frost’s SPECIFIC question work when his broad ones failed?
Reveal answer
A broad question lets the source drift and evade; a specific one names the exact claim, leaving no room to move. Specificity constrains the answer into the open.
- 2
Loftus changed only the verb (“smashed” vs “hit”) and witnesses “remembered” differently. Which test does that question fail?
Reveal answer
Clean vs loaded — “smashed” smuggles in an assumption of high speed, contaminating the answer rather than retrieving it.
- 3
You ask an AI “is social media bad for teenagers?” and get a long hedge. Rewrite it to pass all three tests.
Reveal answer
Make it open + specific + clean, e.g. “What are three documented effects of daily Instagram use on 14–16-year-olds’ sleep, with sources?” — no yes/no, a named target, no smuggled verdict.
How to cut a question sharp
The Sharp-Question Test
Three tests you run on any question before you ask it — whether the source is a witness, an interviewee, or an AI. Sharp = open AND specific AND clean.
**Sharp:** 2 and 4 — both name a population, a condition, and a measurable, falsifiable outcome. **Weak:** - **1 — loaded.** "Obviously" and "lazy" smuggle the conclusion into the question. - **3 — compound + vague.** Two questions (education *and* economy) and no definition of "good/bad" — cannot be answered, only emoted. Sharpened #1: *"Do students who draft essays with AI show lower un-aided writing scores after one term?"* — a claim that could turn out false.
Sharpness = open + specific + clean, stacked. 0 = fails two or three tests (closed+vague, or loaded). 1 = passes one (open-but-vague, or specific-but-closed). 2 = open+specific+clean. 3 = adds a checkable constraint (format/source). The loaded "ruined art" question scores 0 no matter how specific — a smuggled verdict poisons the whole question.
Build Bank
Sharp questions are how you interrogate AI output as a source instead of trusting it as an authority — the core literacy your portfolio is built to prove.
Interrogate a source — and prove your questions are sharp.
Pick a source you actually need to get a real answer out of: an AI model, a person you could interview, or a topic you'd search. Write three questions you'd use to get usable answers — then, for each, run the Sharp-Question Test on your own wording:
- The question, written out in full.
- Open or closed? (and why that's right for what you want)
- Specific or vague? (name what it pins down)
- Clean or loaded? (what assumption are you NOT smuggling in)
- Your prediction: what kind of answer will this pull?
You pass when at least two of your three questions are open, specific, and clean at once — and when you can name a weak question you rejected and say which test it failed.
- ✓ [
- ✓ ]