Frame the Hypothesis

AI makes prototyping so cheap it hides the question that actually matters: is the problem real? A working demo proves nothing about demand. This skill turns "I have an AI idea" into "here is the one assumption that decides whether this is real, and the cheapest experiment that could prove me wrong", and it forces the model to attack the idea instead of flattering it.

The method

Sharpen the problem, score the riskiest assumption, write the falsifiable bet, then make the model attack it. Full anatomy, the uncertainty x fatality rubric, the devil's-advocate protocol, and a worked example: references/hypothesis-craft.md.

Step 1: Sharpen the problem, not the solution

Reframe in the customer's words: who exactly ([ICP]) fails to get what painful job done ([PROBLEM]), how do they cope today, and how often does it hurt? Output a one-paragraph problem statement.

Step 2: Name the riskiest assumption

List the beliefs the idea rests on, then flag the one that is both most uncertain and most fatal if wrong ([RISKIEST_ASSUMPTION]). Usually problem or demand risk, not technical risk.

Step 3: Write the falsifiable hypothesis

State it as: "We believe [X]; we'll know we're wrong if [bar]." Commit to the pass/fail bar before running anything. If no result could fail it, it's theatre.

Step 4: Run the structured devil's advocate

The model will confirm your idea if you let it (sycophancy). Instead, make it refute you: "Give five specific reasons this fails, and the cheapest test that would expose each." Dispatch the devils-advocate agent for a harder pass.

Output

• Problem statement (customer's words) + the assumption stack, riskiest one flagged.
• One falsifiable hypothesis with an explicit pass/fail bar and decision rule.
• The five strongest refutations, and which to test first. Next skill: map-the-terrain or architect-before-code.

Constraints

• Adversarial by default: if the model only justifies the idea, the prompt failed. Make it attack.
• Validate problem and demand before solution. Building is not validating.
• Don't invent market numbers; label any figure an assumption to confirm.
• Stay theme-agnostic; the founder supplies the domain.

Dictionary

sycophancy · the 40% rule · the Remove-the-AI test

Copy-paste version

For non-coders: paste into any chatbot (Claude.ai, ChatGPT, Gemini).

Act as a sharp, skeptical AI-native validation coach. My idea: [ONE_LINER]. For: [ICP]. Problem I
think they have: [PROBLEM]. Domain: [DOMAIN].
1. Rewrite the problem in the customer's words: who exactly, what painful job, how they cope now, how often it hurts.
2. List my assumptions, then pick the ONE most uncertain AND most fatal if wrong.
3. Write it as a falsifiable hypothesis: "We believe X; we'll know we're wrong if [bar]." Set the number before I test.
4. Now turn on me: give five specific reasons this idea fails, and the cheapest experiment that would expose each.
Do not flatter the idea. Make me test the problem before the solution, and don't invent market numbers.