After AI takes over, what job will you be assigned?
This skill scans your file directories, then — in the name of the "AI HR Department" — issues you a Human Placement Notice: which of 16 types you are, and the evidence behind the verdict.
You get back a vertical card image you can post straight to social media.
Note on language: the verdict, the 16 role names and the card are all written in Chinese. The skill is built for a Chinese-speaking audience. Everything below describes how it works; the output itself will not be in English.
The tidier you are, the easier you are to replace.
The moment you organize your work into something clean and legible, you have already done the hardest part of the job for the AI. Conversely: directories with chaotic naming, runaway nesting, and structure only you can parse — those the AI cannot learn.
So in this system, mess is a moat.
This is the most important thing about this project, and the reason it can be open source.
The collector — plugins/ai-hr/scripts/scan.py — is ~120 lines of pure Python standard library. You can read all of it. The JSON it emits has no field that holds a filename; only counts, file extensions, and fixed enum keys.
Which means: this is not "we promise not to look at your filenames." It is that the model never gets the chance to see one. Privacy here is a property of the output schema, not a pledge in a policy page.
The skill file also hard-codes the rule: the only permitted way to learn anything about your filesystem is this script — listing directories directly is forbidden.
The script is read-only. It never writes, moves, or deletes anything.
The full argument — why privacy promises fail, what makes one structurally checkable, and how this repo satisfies it, with line-level citations into scan.py — is in docs/privacy-by-structure.md.
scan.py is short enough that any competent agent can read all of it before installing. When we ran the install through WorkBuddy, it audited the code unprompted and reported back: three read-only calls (os.listdir / os.walk / os.stat), no writes, no network requests, no subprocess / eval / exec; the skill file contains no prompt injection and no exfiltration instructions. Its own phrasing for the schema finding was that privacy here is structural, not a promise.
Ask yours to do the same before you install. That is the point of keeping the collector this small.
Requires WorkBuddy or another desktop agent compatible with the OpenClaw skill format.
Simplest way — just ask your agent to do it. Paste this:
请帮我安装一下这个项目 https://github.com/GiaSip/ai-hr ，然后帮我试一下
That is verbatim the message we tested with. The agent read the repo, audited the code on its own initiative, installed the skill, and ran it — one message, no menus, no marketplace dialog.
Or register it as a plugin marketplace (this is what gets you update notifications): open WorkBuddy's add marketplace dialog and enter GiaSip/ai-hr — the full URL works too — then install the ai-hr plugin.
Either way, trigger it with:
给我的电脑画个像
No runtime setup needed — WorkBuddy ships with its own Python.
Four binary axes, combined into 16 types, each with a four-letter code:
| Axis | Meaning |
|---|---|
| R / C | Structure: regular ↔ chaotic |
| M / K | Output: maker ↔ keeper |
| S / B | Rhythm: steady ↔ bursty |
| A / F | Time: deep archive ↔ travelling light |
For example RMBA「节点爆破手」(Deadline Detonator): dormant for weeks, then at every deadline pulls templates out of the archive and assembles something immaculate — a plant that flowers only three days before the due date.
Four of the sixteen are marked rare. Rarity is defined by contradictory trait combinations, not by an extreme value on any single axis.
The verdicts are meant to sting. Social sharing runs on a little offense — something too gentle gives nobody a reason to repost it.
But there is a line: it judges how you pile up files, never who you are. Nothing about appearance, gender, region, or income. The goal is that you laugh, swear once, screenshot it, and post it — not that you actually feel bad.
Every verdict is framed as a pre-assessment: it reflects what a future AI would want in a hire, not a judgment of who you are today.
MIT