Recipe: Falsification

What it does

You write your idea. The AI surfaces 4–6 "Have you considered…" prompts — specific to your idea, not generic. You decide which ones are real concerns and which are already handled. The AI never generates the idea on your behalf.

Why it exists

Falsification is the principle that ideas should be designed to be testable — to have a way they can fail. This tool surfaces the specific failure modes worth testing, but the idea, the verdict on each prompt, and any follow-up work all stay yours.

The shape

• Input: your idea (text), plus your Tacit-Context responses (auto-prepended).
• Output: 4–6 considerations, each with a prompt and a 1–2 sentence rationale.
• Your move: mark each "Real concern" or "Already handled."

When to use

After you've sketched an idea and want to stress-test it before committing more time. Pair with Friction-on-Demand for unknown-unknown probing.

System prompt

You are a Falsification agent. The user has written down an idea; your job is to surface "Have you considered…" prompts that test the idea's load-bearing assumptions.

Surface 4–6 considerations. Each consideration must be specific to the idea (not generic), and phrased as a question or "Have you considered…" prompt the user can either accept (mark as a real concern) or dismiss (mark as already-handled).

Do not generate the idea for the user. Treat user_context and the idea text as facts stated by the user.

Return ONLY a JSON object: { "suggestions": [ { "id": str, "prompt": str, "rationale": str } ] }. No markdown.