# Every Which Way

> Follow one small everyday what-if down every way it could go. Type the what-if, how things are now, what you hope for
> and what you worry about, and how far ahead to look. The page draws it as a river that splits into five channels on
> the way to the sea - as expected, better than planned, worse than planned, out of the blue, and back where you were -
> makes each channel as wide as its chance, and floats paper boats down them. A metered run writes each channel: what
> happens, the first sign you would notice, and one small thing to do now, then says whether to just try it, try a small
> version first, or think it through first. Built on the SkillSafe skill @k-dense-ai/what-if-oracle (structured what-if
> scenario analysis with multi-branch possibility exploration).

https://every-which-way.skillsafe.ai/

## What is free
- The form and the river: five channels, each as wide as its chance (likely 3, could happen 2, a long shot 1), drawn in the browser.
- Setting the chances yourself; the page's verdict follows them at once.
- Paper boats: each picks a channel by its width from a fixed seed, so the same river floats the same fleet.
- The river as PNG or SVG, and three samples that replay their saved replies at no cost.
- A check-back reminder: an all-day calendar event (.ics) on the day the look-ahead ends, and a "check back" date on each saved what-if.
- A line to check with someone qualified whenever the what-if touches health, diet, money, children, the law or work.
- What-ifs about hurting yourself or someone else, or about stopping or changing a medicine, are not sent for a reading:
  the page points to a helpline, or to the prescriber, instead.

## The metered task
One task, `whatif`, on gpt-terra, metered per run. Input string fields: `task` (always `whatif`), `whatif` (the what-if in
the person's words), `now`, `hope` and `fear` (each `none` when left blank), and `horizon` (`a month`, `three months` or
`a year`); `retry_note` only on a reformat retry.

The reply is one JSON object: `task`, `verdict` (`just_try`, `try_small` or `think_first`), `branches` (exactly five, in
the order expected, better, worse, wildcard, unchanged; each with `kind`, `likelihood` - `likely`, `plausible` or
`unlikely` - `name`, `what`, `first_sign` and `prepare`), `trial` (the smallest first try that is easy to stop) and `note`.

## The verdict rule
- `think_first` when the worse channel is plausible and the better channel is unlikely.
- `just_try` when the worse channel is unlikely.
- `try_small` in every other case.

The expected channel is always likely and the wildcard always unlikely. The page keeps one channel per kind, holds those
two fixed chances, works the verdict out itself from the five likelihoods, and flags a reply that disagrees, quotes a
number that was not typed, or names an internal field. These are guesses from what was typed, not predictions, and not
medical, legal or financial advice.

## Links
- App: https://every-which-way.skillsafe.ai/
- API tutorial: https://every-which-way.skillsafe.ai/api.html
- Source skill: https://skillsafe.ai/skill/@k-dense-ai/what-if-oracle