mingli-master asks pip to break system packages to install one chart library

The whole skill is one markdown file and four scripts

The README is written in Chinese throughout, and the deliverable it describes is not an application. It is a skill directory: `SKILL.md` at the root, a `scripts/` folder holding `calculate_chart.py`, `gezhi_rules.py`, `generate_html.py` and `calibration.py`, a `templates/chart_template.html` for the output, and `references/` for the interpretation guide, the fourteen major stars and the four transformations.

The install paths make the shape explicit. Claude Code reads `~/.claude/skills/mingli-master/`, Codex CLI reads `~/.codex/skills/mingli-master/`, and Cursor reads `~/.cursor/skills/mingli-master/`. A runtime that has no Agent Skills support can still use it by pasting the contents of `SKILL.md` into a conversation, because the file is markdown with YAML frontmatter and nothing else.

The repository root also holds a `.DS_Store`, the macOS directory listing artefact, and there is no `requirements.txt`, no `pyproject.toml` and no test directory. The one dependency is therefore a prose instruction rather than something a tool can check.

The install line ends in break-system-packages

The single prerequisite is iztro-py at version 0.5.0 or newer, and the command given for it is:

python3 -m pip install iztro-py --user --break-system-packages

`--break-system-packages` is the part worth stopping on. It is the flag that tells pip to write into an interpreter the operating system has marked as externally managed, which is the guard that stops a package install from breaking the distribution's own Python. `--user` puts the package in the user site directory, so the blast radius is smaller than the flag implies, but the flag itself is what disables the check.

The version floor is stated in prose and repeated nowhere else, so nothing enforces it. iztro-py is the pure Python port of iztro, the JavaScript chart library credited in the acknowledgements with more than 4000 stars, and the credit for the V2 pattern system goes to a community contributor rather than to the author. If the package resolves to an older version, the failure is a rule engine disagreeing with a chart rather than a clear error.

Three runtimes are named and one clone command is written

There are three ways to get the skill onto a machine and only one of them is a command you can copy.

npx skills add learnwithu/mingli-master

git clone https://github.com/learnwithu/mingli-master ~/.claude/skills/mingli-master/

The npx line goes through a third party installer and the clone line targets the Claude Code directory. The install table names Codex CLI and Cursor with their own directories, but no second clone command is written for either, so those two rows have a path and nothing else. The recommended route is neither of these: it is to open the agent and ask it to install the skill from a URL, which puts a model between you and a directory copy.

The documentation also notes that the project follows the open Agent Skills protocol and therefore runs in any skills-compatible agent runtime, while the rest of the page is written for the three runtimes in the table.

The pattern judgement is a rule file with published coefficients

The interesting part is that two of the five steps are code and only one is the model. `calculate_chart.py` calls iztro-py for the twelve palace star placement, the flying transformations and the decade limits. `gezhi_rules.py` then does three fixed jobs.

The first is empty palace borrowing: an empty palace takes stars from its opposite palace, tagged with an energy coefficient of 0.7, dropping to 0.5 when the opposite palace carries a malefic star, and the borrowed palaces are marked with a dashed line in the chart. The second is pattern judgement across the three-and-four corners, recognising six named patterns including 杀破狼, 紫府武相, 机月同梁, 阳梁昌禄, 巨日同宫 and 马头带剑, each scored on three axes. The third scans the year, decade and birth-year taboos for overlap and turns that into a pressure score, where 2.5 or more is high risk and 1.5 is a warning. The year is chosen with a `--year` parameter that defaults to the current year.

The model's job is the prose, and the page is explicit that given the same birth data the verdict is always the same.

V2 shipped a correction to a table that mis-scored whole years

The single release, v2.0, is described as a deterministic rules engine, two themes and three rule corrections, and the corrections are more interesting than the features.

The first one is a wrong star table. The 甲, 乙 and 丙 entries in the four transformation taboos had been assigned the transformation stars instead of the correct ones, that is 甲廉贞, 乙天机 and 丙天同 where the standard reading is 甲太阳, 乙太阴 and 丙廉贞. Because that table feeds the pressure audit, the error shifted the whole reading for any affected year, and it was found while merging a community contributed rule set and cross-checking it. The fix was verified by charting three separate birth years with the iztro engine.

The other two are narrower: the year pillar calculation had been missing the earthly branch and was hardcoded, with 2026 listed as a 丙午 year, and entering a lunar date left the Gregorian date empty, which made the year audit unusable for that input path. A fourth change, extending the 杀破狼 pattern to a full three-and-four-corners match including borrowed stars, is listed among the corrections rather than the additions.

The structure listing stops in the middle of a filename

The repository structure section of the README prints the tree and then ends partway through the last entry, at `four_hua_re` under `references/`. The three names above it are complete: `interpretation_guide.md` for the writing style, `stars_reference.md` for the fourteen major stars plus the six auspicious and six malefic stars, and a partially named file that the acknowledgements call the four transformation rules.

Two other behaviours are described rather than listed. The HTML output carries two themes, a gold and black classical one and a glowing card one, switchable from the top right of the page, with the choice remembered in browser local storage. And the palmistry cross check is optional: if you supply a palm photo or a written description, the skill extracts features and marks where the palms and the chart agree and where they conflict.

The chart itself is said to survive an empty interpretation. If `reading.json` is empty the generator still produces the full skeleton chart, with the dashboard and the pressure table rendered from data alone. The worked example on the page is a chart for Elon Musk, with the rendered image kept at `assets/musk-mingpan.jpg`.

One tag, and a stated promise of no cadence

The repository has one release, v2.0, dated the same day as its last commit on 2026-08-15. There is no 1.x history behind it in the tags, so there is nothing to compare the current rule engine against and no upgrade path described between versions.

The maintenance position is stated rather than implied. A line near the licence says the project is maintained out of interest and makes no promise about update frequency, which matches a repository whose last push and only tag share one date. The licence section gives MIT and says in Chinese that you can use it and change it as you like, and the credits name the two people and the two chart libraries the project builds on.

The page closes on a line about the practice itself rather than the software, saying that fortune reading is a conversation and not a performance, and that accuracy approaches as the conversation goes on. That is a claim about the reading, not about the chart, and nothing in the repository measures it.

Editorial conclusion

mingli-master is worth reading as an example of a skill that moves arithmetic out of the model and into code, because the placement is computed by iztro-py and the pattern judgement is a rule file rather than a prompt. Two things to check before using it. The install line ends in `python3 -m pip install iztro-py --user --break-system-packages`, which removes the guard your operating system puts on writing into the system interpreter, so read the version floor before running it. And the astrology itself carries no verification: the charts are arithmetically consistent, the reading is written by a language model, and the page closing line says the accuracy approaches through conversation rather than promising a result.

Frequently asked questions

Official sources

1. Issues
2. learnwithu/mingli-master on GitHub
3. License: MIT
4. README
5. Releases