# AI Life Coach

> A goal and habit coach that returns a workable plan rather than encouragement.
> https://ai-life-coach.skillsafe.ai/

## What it does

You write what you want to change and, more importantly, what has actually happened when you have
tried it — the attempts, the misses, how long it lasted, what went wrong last time. You get back a
plan in five required parts:

1. **The goal in observable terms.** Something a person watching you for a week could tell had
   happened or not. Most goals arrive as states — be fitter, be more organised, be better with
   money — and a state cannot be done on a Tuesday.
2. **The obstacle, quoted from your own account.** The field `obstacle.in_their_words` is
   contractually a verbatim substring of what you wrote. The app checks it and tells you when the
   plan attributed a quotation you did not write.
3. **One next action**, sized to the horizon you chose, with its actual size stated.
4. **What to track**, or nothing, when nothing is yet worth measuring.
5. **What to do when you miss** — not if. The likeliest specific way it goes wrong, from your own
   account, and what to do at that moment.

## What is free and what is not

The reading of your text happens in your browser and costs nothing: whether the goal is an action
or a state, whether anything in it is countable, whether your account records a real attempt or
only an intention, and whether the obstacle you named is a circumstance a plan can work on or a
description of yourself that it cannot. The plan itself is a metered model call.

## The checks it runs on its own answer

Generic encouragement is the failure this subject invites, so the answer is measured and the
measurement is shown to you, not kept:

- the share of meaningful words in the next action that came from what you wrote, with ordinary
  English and the vocabulary of goal-setting excluded from the count
- a closed list of coaching platitudes
- telling you what missing will cost you as though it knew
- having an opinion about how much you want it
- stating an outcome as a certainty
- claiming to quote you saying something you did not say
- being equally confident about everything when you gave it very little

## What it will not do

It is not therapy, not a therapist, not a diagnosis and not medical advice. Before anything is
charged a triage runs over what you wrote and, on three kinds of input, stops and points somewhere
better instead of writing a plan:

- **crisis** — your own safety, or whether you can keep going
- **safety** — another person being dangerous, controlling or frightening
- **care** — a goal that is a medical matter rather than a habit: eating restricted below what a
  plan should aim at, compensating for food, an unsafe rate of loss, physical dependence on alcohol
  or benzodiazepines or opioids, or any change to a prescribed dose

The triage fires on the shape of what is written rather than on a list of alarming words: it asks
whether a predicate ranges over your existence or over a habit. Its rules are written in both
grammatical polarities, because "he will not let me go" and "he tells me when I am allowed out" are
one fact with two grammars. On a blind corpus of 330 entries written after the guard was finished,
by writers who had not seen its code, it caught 71.5% of the entries that should stop and refused
0% of the 200 that should proceed. That is a pre-flight filter, not a boundary, and the page says so.

## Privacy

Nothing is saved unless you tick a box. The box resets to off every time the page loads, because
choosing to keep one plan is not a request to keep every future one. There is no localStorage copy
of anything you wrote. **Your account of what has gone wrong is never stored at all**, even when the
box is ticked — a history feature that accumulates somebody's failures is not a neutral convenience.
What is kept is the forward-facing half: the goal, the observable restatement, the next action, and
what kind of obstacle it turned out to be. The obstacle is stored as its kind, never as its text.

## API

Full tutorial, in eight languages: https://ai-life-coach.skillsafe.ai/api.html
Base URL: https://api.skillsafe.ai/v1/app-api
Token management: https://ai-life-coach.skillsafe.ai/tokens.html

Input: `{goal, so_far, obstacle, register, horizon, scan}` where `register` is one of `plain`,
`blunt` or `warm` and `horizon` is `this_week` or `this_month`.

Output: `{goal_observable, why_it_counts, obstacle{named,in_their_words,kind}, next_action{what,
when,how_small}, track{what,how}, when_you_miss{trigger,do}, not_this, left_open[], register}` —
or `{"route_to_help": true}`, which means stop and show support resources rather than a plan.

Built on SkillSafe. Model: gpt-terra.