Make AI ask first, before it guesses

You hand an AI a half-formed request. It doesn't push back — it picks the most likely reading and runs. Half the time the guess is fine; the other half you spend twenty follow-up replies steering it back. This makes the model ask first.

The trick is one instruction: tell the model to turn the ambiguity it would otherwise hide into a short questionnaire. It returns a JSON object; you paste it into passbackai.com, which renders it as a form — radio buttons, checkboxes, an "Other" field. You answer the decisions in 90 seconds, Copy the bundle, and paste it back. Now it proceeds on your answers, not its assumptions.

The prompt

Drop this into Claude, ChatGPT, or Gemini. It tells the model to interview you, and to emit the questions in the exact JSON shape passbackai.com renders:

I want you to understand me better and avoid back-and-forth guessing. Please review our conversation and ask me clarifying questions to resolve any ambiguities.

To do this, turn the unresolved issues into an PassbackAI JSON questionnaire. You must STRICTLY follow this exact JSON schema:

{
  "version": "1",
  "skill_version": "1.2",
  "title": "Clarification Needed",
  "routing": { "from": "Assistant", "return_prompt": "Please paste your answers back here." },
  "questions": [
    {
      "id": "q1",
      "section": "Theme of the question",
      "question": "The decision I need you to make",
      "context": "A sharp sentence explaining why this is ambiguous or blocking our progress",
      "multi": false,
      "options": ["Concrete Option 1", "Concrete Option 2", "Concrete Option 3"],
      "open_field": { "label": "Other:", "placeholder": "Specify here..." }
    }
  ]
}

RULES:
1. Output the valid JSON STRICTLY inside a markdown code block. Do not miss any brackets or quotes.
2. Keep it to the decisions that actually block progress — at most 12 questions.
3. Give 3–4 concrete, mutually distinct options per question. Never synonyms.
4. Use "multi": true only when more than one answer is genuinely valid.

Copy it, paste it under whatever you were discussing, and send. The model returns a JSON block. Copy that block, open passbackai.com, and paste — the form opens.

Why a form beats a wall of questions

Ask a model "what's unclear here?" and you get back a list of paragraphs. You read three, you nod, you close the tab — none of them ends in a button you have to push, so the deciding never happens. A form is different: each question ends in radio buttons you can't skim past. You either pick an option or you mark "I need to check with: legal" — both are answers. The decisions you can't make become visible as someone else's, and you forward them as one bundle instead of seven Slacks.

Answer it yourself, or send it on

The same JSON works two ways. By default the model emits a form for you — you open PassbackAI, answer, Copy the bundle, paste it back. Or it fills in the routing.from field with whoever's waiting ("Alex sent this") and a return_prompt telling the answerer how to send it back, so you can forward the form to the person who actually holds the decision. They answer on PassbackAI in 90 seconds and DM you the bundle.

Want it automatic?

If you'd rather not paste a prompt every time, install the Claude skill — the protocol above is built in, so a phrase like "interview me" or "turn this into a questionnaire" fires it. One-time setup, no API key, nothing leaves your machine.

PassbackAI points an arrow at the answer — paste a reply, mark every wrong line, send the fixes back in one paste. This points the arrow back through the question — make the model surface what it doesn't know before it spends a thousand tokens guessing. Same diagnosis, opposite direction. Both end on the same form.