Real charts. AI prose. That's the rule, and we mean both halves of it.
AstroGuide uses large language models to turn structured astrological data into readable prose. We want to tell you exactly what that means — and what it doesn't — so you can decide how much to trust the words you're reading.
What the AI does
After the math is finished, a language model receives the structured result and writes English (or Russian, or Spanish) prose around it. Its job is exclusively the prose layer:
- Find a good first sentence for your daily forecast.
- Phrase the meaning of "Saturn squares your natal Moon" without the doom of a hundred old astrology books.
- Vary tone across day-week-month so they don't read identically.
- Translate the same prose for the locale you've chosen.
That's it. The astrology itself — which planet is where, which aspect at what orb, which house cusp falls in which sign — is computed by Swiss Ephemeris before any AI sees anything.
Guardrails and limits
- It does not calculate the chart. The geometry is fixed before it gets involved.
- It is instructed to use only the supplied aspects, transits, and planetary positions. A generative model can still make a mistake, so calculated chart data remains the source of truth.
- It is instructed not to predict specific future events. No "you will meet someone tall on Thursday".
- It is instructed not to give medical, mental-health, financial, or legal advice. High-risk and crisis wording is intercepted before generation, but no automated safeguard is infallible.
- It does not receive your account email, Google account ID, purchase token, notification token, or journal entries.
- It is not a professional adviser. Generated text can be wrong and must not replace medical, mental-health, financial, or legal advice.
Models we use
We route through OpenRouter to specific models, picked for the job:
- Claude Haiku 4.5 — the configured primary model for most readings.
- Gemini 2.5 Flash — the configured fallback if the primary model is unavailable.
- Claude Sonnet 4.6 — used for selected longer readings such as tarot spreads.
We choose specific frozen versions. If we change models, we'll note it here.
Each request asks OpenRouter to use zero-data-retention routing and deny provider data collection. Availability still depends on the selected provider supporting that privacy mode.
What data it sees
Before processing begins: when you tap an action that requests generated content, AstroGuide sends the relevant data below over HTTPS to its backend, then to OpenRouter and the selected model provider. This transfer happens before the generated text can be returned. If you do not request an AI feature, no AI request is made.
The prompt sent to the model contains, roughly:
- Your computed astrological positions (planet, sign, degree, house).
- The aspects between them, with orbs.
- Current transits for the period being asked about.
- The locale of the requested output (en / ru / es).
- The reading type (daily, weekly, monthly, natal, synastry, tarot).
- Depending on the feature: your display name, age derived from birth date, birth date/time/place, the approximate current city or town you typed (not GPS), gender, a partner profile, or the question you typed.
The prompt does not contain your account email, Google account ID, purchase token, notification token, journal entries, or analytics history. See the Privacy Policy for processors and retention.
When it gets it wrong
Language models occasionally write nonsense, contradict themselves, or produce harmful text. Use the report action next to generated content in the app. Choose a reason and, optionally, add context; the report is sent to AstroGuide for review without opening another app.
We don't claim the prose is perfect. We claim it's honest about what it is.
AstroGuide currently has no non-AI replacement for generated readings. You can avoid AI features and still use local journal, calendar, settings, and other non-generated parts of the app.