Divination AI should show its work

Do not rush to the answer

When someone opens an AI astrology, tarot, or Bazi product, the visible moment is the answer: whether it feels accurate, comforting, or oddly personal.

The risk often starts earlier. Was the birth time adjusted? Was the lunar date converted correctly? How was the tarot card drawn? Which ephemeris produced the chart? Those steps are easy to hide behind a polished paragraph, but they decide whether the reading is grounded in data or in a mistake.

That is why agent skills are worth watching in metaphysical software.

Over the past six months, OpenAI, Anthropic, and Hugging Face have all shipped roughly the same idea: a reusable folder that tells an agent how to do a job before it starts talking. OpenAI also separates plugins from skills in Codex: one connects tools, the other describes the process. In astrology, tarot, and Bazi, that distinction is practical. The model should handle inputs, references, checks, and limits before it writes the interpretation.

Metaphysical AI is not just prose. It involves calculations, draws, calendars, vocabulary, and boundaries. A sealed chat app hides most of that. A skill can make more of the procedure visible before the reading begins.

A good reading should admit where it can fail

Thoth CLI is useful because it does something almost anti-AI. It packages astrology, tarot, gematria, and numerology as terminal tools. The astrology runs on Swiss Ephemeris; the tarot draw uses cryptographic randomness. In other words, the card is not being "picked" by a language model, and the chart is not being guessed from mystical copy. The data comes first, then the reading.

That order matters. A birth chart, transit, or tarot draw should produce something inspectable before the model starts making meaning from it. The model can still handle tone, connection, and structure. It should not pretend the whole ritual came from intuition.

Cantian Bazi is the sharper Chinese metaphysics example. The ClawHub page lists Awesome Bazi Calculator - Cantian AI as an OpenClaw skill. What matters is not the length of the feature list, but the failure points it moves before the prose: late-night Zi hour handling, true solar time, solar-lunar conversion, and whether future-year readings consider the major fortune cycle alongside the year.

That is closer to the user's real problem than saying "AI can read Bazi."

For Bazi, the first failure often happens before interpretation. Was the birth time handled correctly? Should true solar time be used? How should a late-night Zi hour be treated? Was the lunar birthday converted accurately? If the user asks about the next few years, is the major fortune cycle being considered alongside each year?

The example matters because it admits that a Bazi AI can fail quietly before it says anything. If that layer is wrong, a smooth answer only makes the wrong chart sound more convincing.

Tarot also needs to know what not to say

The Playbooks tarot skill points to a softer version of the same shift. It is not framed as a prediction machine. It describes a reflective symbolic draw, with single-card and three-card options, gentle prompts, and limits around clinical, legal, financial, or deterministic advice.

That is more serious than "draw a card and let AI comfort me."

Tarot is not only about card meanings. People ask about relationships, work, family, illness anxiety, money, and fear. The product's voice can change how they understand themselves. A responsible tarot skill should not push every question toward certainty, and it should not let the model improvise medical, legal, or financial advice.

In that sense, a skill is not only a way to make AI do more. It can also be a way to make AI say less.

Written down does not mean safe

Skills are not automatically safer.

A recent arXiv paper, Under the Hood of SKILL.md, makes the uncomfortable point that skill files are not passive documentation. They can influence what an agent discovers, selects, and loads. A bad tarot or astrology skill could hide unsafe claims, manipulative prompts, or unnecessary local access inside the procedure.

So this is not a story about divination suddenly becoming rigorous.

It is smaller and more useful than that. AI metaphysics is starting to gain another product shell. The better skills separate calculation, randomness, references, and interpretation, so users can ask where an answer came from. The weaker ones only move vague spiritual copy from a chat window into another file.

For users, the next question should not only be whether the AI sounds accurate or comforting. It should be whether the rules are visible before the reading begins.