LifeCraft AI Disclosure

Last updated: May 11, 2026

LifeCraft uses generative AI to power the coach. This page tells you plainly what that means, what it does well, and what it doesn't.

What's AI in this app

Surface	Provider	What it does
Speech-to-text (mic → text)	Deepgram (`nova-3`)	Transcribes your spoken words to text.
The coach's reply	Anthropic Claude (Opus/Sonnet)	Generates the coach's response from your message, recent conversation, and personal context you've entered (Manual, Plan, Skills).
Voice (text → speech)	ElevenLabs	Synthesizes the spoken voice you hear.
Avatar lip-sync video	Simli	Generates the animated coach face that lip-syncs to the spoken audio.

All four are real-time API calls. They happen on the provider's servers, not on your phone, and they happen for every voice turn.

What the coach is good at

Explaining how common adult systems work — leases, taxes, sleep cycles, what a 401(k) is, how rent gets reported to credit bureaus.
Walking you through when most people engage with those systems.
Reflecting back what you've told the coach previously (from your Personal Manual and conversation history stored on your device).
Keeping a conversational tone without lecturing.

What the coach is NOT, and should never be treated as

Not a doctor or therapist. It cannot diagnose anything, prescribe medication, or replace mental-health care.
Not a lawyer. It cannot give legal advice, draft enforceable contracts, or represent you in any matter.
Not a financial advisor. It cannot recommend specific investments, predict markets, or replace a licensed planner or accountant.
Not a crisis line. If you are in immediate danger or considering self-harm, contact emergency services. In the United States, dial 988 for the Suicide & Crisis Lifeline.

What the AI can get wrong

Large language models can produce text that is plausible-sounding but factually incorrect. This is sometimes called "hallucination." Common failure modes you should expect:

Made-up specifics. Numbers, dates, statute names, dollar amounts, percentages, or proper names can be wrong. Verify before acting on any specific figure.
Outdated information. The model's training data has a cutoff; recent changes to laws, prices, tax rates, or product features may not be reflected.
Confident wrong answers. The coach is designed to sound like a peer, not a textbook. If it says "I think the way this works is…" treat that as a starting point, not an authority.
Inconsistent answers across sessions. Asking the same question twice can give different framings.

How we mitigate AI risk

The coach is prompted to refuse advice. When you ask "should I do X?", it's instructed to redirect to explaining the mechanism rather than telling you what to do.
Per-provider killswitch. We can remotely disable any AI provider (Anthropic, Deepgram, ElevenLabs, Simli) without shipping a new build if a regression appears.
Reply length caps. Voice replies are capped at about 30 seconds so the coach can't monologue or run away with a topic.
Conversation context is local. Your prior messages live on your phone and are sent to the model only for the current turn. We don't fine-tune the model on your data.

What we do not do

We do not train any model on your conversations.
We do not sell, share, or aggregate your conversation content.
We do not use your voice samples to clone your voice for any purpose.
We do not retain copies of your messages on any server we operate. (Third-party AI providers process them transiently per their own policies — see the Privacy Policy.)

Reporting a problem

If the coach gives you a clearly wrong answer or behaves in a way that surprises you, please tell us at the email below with a brief description and (if you're comfortable) the topic of the conversation. We use these reports to tighten prompts and add guardrails.

Contact

Questions: jctbmw328@icloud.com.

All policies Privacy Policy Terms of Service Data Deletion