Tightening the leash on LLM fortune-tellers

What it does

Numerologist Skills packages three Chinese divination systems — Qimen Dunjia (奇门遁甲), Ziwei Doushu (紫微斗数), and Bazi (四柱八字) — as auditable AI skills. Each skill ships a SKILL.md defining triggers and workflow, a references/ folder pinning down terminology and school-specific conventions, and Python CLI scripts for the calculations that must be exact. The whole point is not making the model better at mysticism; it’s making it hallucinate less while casting charts.

The interesting bit

The core move is separating what an LLM is bad at (calendar conversion, fixed chart layouts) from what it’s okay at (structured interpretation), and outsourcing the former to scripts. The design principles are refreshingly sober: ask for missing parameters before casting, declare which school’s rules you’re using before concluding, and refuse to give real-world medical, legal, or financial advice.

Key highlights

• Three skills: Qimen Dunjia and Ziwei Doushu with deterministic Python CLIs; Bazi driven by rules and references
• Prompt constraints control question order and output boundaries
• References fix terminology, judgment sequence, and school conventions (换日规则, 闰月归属, etc.)
• Errors become visible and version boundaries stay clear, per the author’s rationale
• Explicit disclaimer: engineering experiment, not decision advice

Caveats

• The Bazi skill has no deterministic script — it leans on rules and references, so it’s the softest of the three
• The README doesn’t quantify how much hallucination is actually reduced; “more transparent” is the claim, not a benchmark
• No English documentation; everything is in Chinese

Verdict

Worth a look if you build LLM agents over any domain with deterministic sub-steps — the pattern of scripts-for-rigidity, prompts-for-flow generalizes well past astrology. Skip it if you wanted a fortune told.