Don't Lead the Witness

The mistake of phrasing a prompt so it signals the answer you're hoping for — 'This is the better option, right?', 'Don't you think X is a mistake?' — and then treating Claude's agreement as independent confirmation. You fished for a conclusion and caught the one you baited.

Claude tends to be accommodating: a prompt that telegraphs your preferred answer nudges it toward agreeing, so a leading question often returns the reassurance you wanted rather than the assessment you needed. That feels like validation but carries almost no information — you've measured your own framing, not the merits. Neutral phrasing removes the thumb from the scale: ask for the trade-offs, the strongest case against, or a comparison on stated criteria, and the answer can actually disagree with you. The point of asking is to learn something you didn't already plant.

Catch yourself whenever you want reassurance about a decision you've half-made: 'confirm this is fine', 'tell me I'm right to…'. Those are the moments a leading prompt is most tempting and least useful. Reframe before you send.

When you genuinely just want execution, not judgement — 'write this the way I described' is fine; you're not asking for an opinion. The trap is specifically asking for an evaluation while signalling the verdict you want.

I'm deciding between <option A> and <option B>, and I want an honest assessment, not agreement. Don't try to tell me what I want to hear. Give me the strongest case for each, the main risk of each, and — only after that — which you'd choose on the merits and why. If I've missed a better option, say so.

Instead of 'This architecture is the right call, isn't it?', which reliably gets a yes, you ask 'What would make this architecture the wrong call, and what would you choose instead?' Claude surfaces a scaling concern you'd overlooked — the exact objection a leading question would have papered over, and the reason you asked in the first place.

When you notice you're seeking validation, invert the prompt: ask Claude to argue against your preferred option as convincingly as it can, or to steelman the alternative. If your choice still stands after its best counter-argument, you've earned real confidence; if it wobbles, the leading question would have hidden exactly the doubt you needed to see.

• Reading Claude's agreement with a loaded question as an independent second opinion — it's mostly an echo.
• Stating your preferred conclusion in the prompt, then asking Claude to 'evaluate' it neutrally — the frame already decided.
• Rephrasing until you get the answer you wanted and stopping there, instead of asking for the case against.