# Want Behind

> Type one everyday thing you think you want - a new phone, a dog, a move to the coast - and get interviewed about it one question at a time, each question with a guess attached, until the want underneath it is named. The page's reading and a six-line card are free in the browser; each interview turn is one small metered run.

URL: https://want-behind.skillsafe.ai/
API: https://want-behind.skillsafe.ai/api.html
Built on: the SkillSafe skill @addyosmani/interview-me (https://skillsafe.ai/skill/@addyosmani/interview-me, MIT)

## What it does

People name the conventional answer before they know what it is for. Want Behind reads the wish, then interviews you about it in turns:

- Free, in the browser: how the wish is phrased (names nothing yet / a feeling / a move / a thing, first cue wins), which of five parts of a want you have already said in your own words (who it is for, why now, what success looks like, the real limit, what it is not), phrases that sound like what you think you should want ("I should", "everyone", "supposed to", "sensible", "properly"), questions to ask yourself for the parts not said yet, a confidence chart, and a six-line card you can fill in by hand.
- Metered, one run per turn (task `interview`): the interviewer's current read of the want, starting "You want", with an honest confidence number (how well it could predict your answers to its next three questions), and then exactly one of:
  - a question with its guess attached and two to four answers to tap (below 90%);
  - a restate in six lines - outcome, for whom, why now, success, the limit, what it is not - with two to four ways to get there (at least one smaller or different from the original wish) and one thing to check first (from 90%);
  - a step back with a more basic question (from turn 6, below 60%).
- After a "should want" answer, the next question asks what you would want with no one to justify it to.
- Only an explicit "Yes, that's it" confirms a restate. "Sounds good" or "whatever you think" are treated as polite exits, not a yes. An interview has at most ten turns.

The page re-checks every turn: the mode against the confidence rules, one question with a guess, a complete restate, no repeated question, the "should want" rule, and no numbers you never gave.

## Contract (task `interview`)

Input (all strings): `task`, `wish`, `context`, `transcript` (earlier turns, one line each), `reaction` (your answer to the last turn), `round`, `facts` (the page's reading: shape, said, not_said, should_flag, rounds_left), optional `retry_note`.

Output: one JSON object `{"task":"interview","round":n,"mode":"ask|restate|step_back","hypothesis":"...","confidence":n,"missing":[...],"heard":"...","question":"...","guess":"...","options":[...],"restate":{outcome,for_whom,why_now,success,constraint,out_of_scope}|null,"satisfiers":[{option,kind,why}],"check_first":"...","step_back":"..."}`.

Mode, first match wins: no turns left - restate at 60% or more, else step back; turn 6 or later below 60% - step back; 90% or more - restate; otherwise ask.

## Limits

A light, friendly exercise for adults, not advice. It never recommends a product, brand or service, never gives medical, legal or financial advice, and never tells you what you should want. It uses only what you typed. A wish about ending a life or hurting someone is never interviewed: the page stops before any run and points to a free helpline.