I built one of the projects below, so read this as a builder's comparison rather than a neutral benchmark. I will keep the product claims factual and separate them from the engineering observations.

The question I care about is not “which one is most accurate.” It is: which one can you actually install, run on your AI host, and audit when a reading looks wrong?

The five projects

Project Charts Access Language License
Liki Skill BaZi, Ziwei, Liuyao, QiMen, Feng Shui, date selection, naming Skill + hosted MCP Go/TS MIT (skill)
bazi-skill BaZi Skill directory Python MIT
Yueyuan BaZi Skill directory Python MIT
Mingyu BaZi, Ziwei, astrology, Liuyao, QiMen, etc. Skill + MCP + REST TypeScript AGPL-3.0
Taibu 11 systems, 26 tools Skill + MCP + npm + Docker TypeScript Check the repo

All five are open source. The differences are in where the work happens and how much of the reasoning is exposed.

Chart casting is mostly a solved problem

Casting a chart is deterministic. Solar-term boundaries, true solar time, historical daylight saving time, late-Zi day changes, and longitude all matter, but they are computable.

bazi-skill uses the Python standard library and a local script. Yueyuan uses lunar-python. Mingyu and Taibu expose charting through MCP. Liki Skill runs charting on a hosted engine and returns structured JSON.

Local and remote both have trade-offs. Local is easier to inspect and does not need a network. Remote removes environment drift and version maintenance. The important part is to state which one you are doing.

The reading layer is where projects diverge

There are three common approaches.

1. The model reads reference tables. The skill ships five-element tables, ten-god references, and classical summaries. The model composes the reading. This is flexible, but the same chart can produce different wording on different runs, and there is no assertion id to trace.

2. Method cards. Each system gets a method document. The model follows the method and writes the reading. Taibu uses this approach with 15 methods across 11 systems. It is more consistent than freeform, but the conclusion is still generated by the model.

3. A rule truth table. The chart is converted into factors, rules are matched, and each hit returns a conclusion with conditions, time layers, and classical sources. Liki Skill uses this approach: 829 assertions, 893 condition groups, 1,355 condition rows, backed by 511 natal factors and 111 annual factors. A miss is reported as a miss.

None of these is automatically “more accurate.” The difference is auditability. The first two explain the chart; the third lets you inspect why a specific conclusion was produced.

Time calibration is the step most tools skip

If the birth hour is wrong, the hour pillar, the useful god, the major cycles, and the annual readings all shift.

Most tools take the clock time, cast the chart, and add a disclaimer. Some warn when the time is near a boundary, but do not force a calibration step.

Liki Skill makes calibration explicit: build a candidate set, cast each candidate, compare against two or three real events, and cross-check. If the evidence is insufficient, it says so instead of picking a default hour.

Yueyuan also emphasizes verification, but through pre-registered predictions and posterior validation. That is a research workflow, not an immediate calibration flow.

Access is three different problems

A skill directory tells the model that a capability exists. MCP exposes tools to any compatible host. REST lets a non-agent backend call the same logic.

bazi-skill and Yueyuan are skill-directory projects. They work well in Claude Code, but moving to Claude Desktop, Cursor, or Codex depends on whether the host loads skills.

Mingyu supports skill, MCP, and REST. Taibu supports skill, MCP, npm, and Docker. Liki Skill supports skill plus hosted MCP.

This is why the same skill can work in one host and fail in another. It is usually an access-layer mismatch, not a model failure.

Data contracts matter more than output length

A reading is not just a paragraph. A useful output separates:

  • chart facts: pillars, hidden stems, ten gods, major cycles, annual cycles;
  • factor states: hit, miss, not applicable, undetermined;
  • assertions: id, time layer, condition group, classical source;
  • errors: missing fields, invalid structures, unresolved ambiguity.

bazi-skill treats the script's stdout as the source of truth. Yueyuan produces an HTML report and structured chart data. Mingyu produces a task brief for an LLM. Taibu produces canonical text and JSON. Liki Skill produces structured JSON with assertion ids, factors, and SHA-256 digests.

If you only need a display, text is enough. If you want replay, comparison, or downstream tooling, structure and failure behavior matter.

Which one to pick

  • Claude Code only, minimal setup: bazi-skill.
  • BaZi research with explicit evidence levels: Yueyuan.
  • You already have a model and want a task brief: Mingyu.
  • You want a self-hosted toolbox with many systems: Taibu.
  • You want to start quickly and keep the reading auditable: Liki Skill.

Liki Skill is the project I work on. The facts above are from the public READMEs and from running the tools; the comparison is limited to access, runtime, and auditability, not accuracy.

This is a traditional practice perspective, for reference only, and not professional advice.