# agency-agents-zh: a Chinese-language library of 277 AI expert role definitions

> agency-agents-zh packages 277 plug-and-play AI expert personas as Markdown role files for Claude Code, Cursor and 18 other tools, adding 64 China-market roles on top of a translated upstream catalog. The value is in the role definitions, not in any runtime.

**jnMetaCode/agency-agents-zh** — 🎭 277 个即插即用的 AI 专家角色 — 支持 Claude Code/Cursor/Copilot 等 20 种工具，覆盖工程/设计/营销/金融等 20 个部门。含 64 个中国市场原创智能体（小红书/抖音/微信/飞书/钉钉/Qt 上位机/机械设计）。搭配编排器 agency-orchestrator，一句话即可让多位专家按 DAG 自动协作。

- Repository: https://github.com/jnMetaCode/agency-agents-zh
- Website: https://ao.aiolaola.com/experts
- Stars: 20,964 · Forks: 3,383
- Language: Shell
- License: MIT
- Published: 2026-09-09 · Updated: 2026-09-09 · Language: en
- Canonical page: https://hysenlabs.com/projects/jnmetacode-agency-agents-zh

## The problem: every team rewrites the same marketing and engineering prompts

Most people using Claude Code or Cursor start from a blank system prompt. They write "you are a senior frontend engineer" once, paste it into a rules file, and then repeat the exercise for the next task. The prompt drifts between machines, nobody reviews it, and the deliverables section is usually missing entirely, so the model decides for itself what "done" looks like. agency-agents-zh attacks that by shipping the prompts as files. The README describes the collection as "277 个即插即用的 AI 专家角色" and is explicit that these are not generic prompt templates: each agent has its own persona, professional process and deliverables. The scope is deliberately wide. The repository layout carries 21 top-level department directories, including engineering, design, marketing, finance, legal, hr, security, gis, spatial-computing, supply-chain and game-development, plus a company directory for executive roles such as CEO, CTO, CMO, COO, CPO and CFO. The intended user is not a model researcher. It is a working team that already has an AI coding tool open and wants a reusable persona for a recurring job.

## What actually ships: Markdown role files, an npm package and a count guard

The mechanism is plain text. Role definitions live in department directories and are consumed by whatever tool you point at them: Claude Code, Cursor, Copilot and 18 others according to the README. There is no server, no daemon and no model call inside this repository. The npm package named agency-agents-zh declares version 1.4.0 and lists the department directories in its files array, so an install delivers the Markdown, not a binary. Two supporting files matter more than they first appear. AGENT-LIST.md and CATALOG.md index what is present, and scripts/check-counts.mjs exists to verify that the advertised counts match the files on disk. The package exposes exactly one script, and its name says what it guards. That is a small but telling design choice: the project treats its own inventory as something that can silently rot, which is what happens when 277 files are edited by contributors and the README keeps quoting an old number. The China-market additions are the differentiating layer. The README states there are 64 original Chinese-market agents covering Xiaohongshu, Douyin, WeChat, Bilibili, Feishu and DingTalk operations, plus cross-border e-commerce, government ToG, medical compliance, Qt industrial HMI, mechanical design and livestock-farming record verification. Those are specific enough that they will either match your workflow or not match it at all.

## Installing agency-agents-zh and pointing Claude Code at one role

The repository is published to npm, so the shortest path is a package install. The command below pulls the role files into node_modules; the package ships the department directories and the index files, not a CLI.

```bash
npm install agency-agents-zh
```

If you would rather keep the files where you can edit them, clone the repository instead. The default branch is main and the licence file sits at the root.

```bash
git clone https://github.com/jnMetaCode/agency-agents-zh.git
cd agency-agents-zh
```

After cloning, the department directories are visible at the top level. Pick one role file from marketing/ or engineering/, read it, and copy its content into the place your tool reads project instructions from. For Claude Code that is the project memory file at the repository root; Cursor reads its own rules directory. The repository does not prescribe a single destination, and the README points readers to the desktop client and the online gallery at ao.aiolaola.com/experts for browsing instead.

The one command the package defines is the count guard. Run it after you add, rename or delete a role file, because the README and the catalog files quote totals.

```bash
npm run check:counts
```

The examples directory is the best starting point for understanding intended usage. It contains files such as workflow-landing-page.md, workflow-startup-mvp.md, workflow-xiaohongshu-launch.md and workflow-with-memory.md, which show roles being combined rather than used one at a time.

## Where the design runs out: no runtime, and orchestration lives elsewhere

The most important limitation is structural. This repository defines roles; it does not execute them. The README's own pitch for multi-agent work, "一句话即可让多位专家按 DAG 自动协作", belongs to a companion project, agency-orchestrator, which is distributed separately as a desktop application for macOS, Windows and Linux. If you install agency-agents-zh expecting a scheduler, you will not find one. You get Markdown and a count-checking script. The second limitation is language and market fit. The catalog is Simplified Chinese, with a Traditional Chinese README and a link to the English upstream project. A team that works in English will spend its time translating personas before it can judge whether they are any good, and the 64 China-market roles are only useful if your channels are actually Xiaohongshu, Douyin, WeChat, Feishu or DingTalk. A third issue is that quality across 277 files cannot be uniform. Some roles, such as the industrial HMI or mechanical design entries, encode narrow domain process; others are closer to a well-written job description. The README does not publish a per-role review status, so the only way to judge a role is to read it. Finally, the README is heavy with sponsor placements and course links, which makes it a poor place to learn what a specific agent does. Use CATALOG.md and AGENT-LIST.md for that instead.

## How it compares with the English upstream and with writing your own prompts

The direct alternative is the upstream project, msitarzewski/agency-agents, which this repository describes as its source. The difference is not just translation. The upstream catalog is English-first; this fork translates 213 upstream agents and adds 64 originals aimed at Chinese platforms and regulated verticals. If your work is in English and your channels are Western, the upstream is the shorter path and you avoid a translation layer. If your work involves Douyin campaign mechanics or a government ToG proposal, the upstream simply does not cover it. The second alternative is writing your own role files, which is what most teams do today. That gives you exact control over deliverables and tone, and it costs you the breadth: nobody on a five-person team is going to author 277 personas, and consistency across them is hard to maintain. The pragmatic middle is to take a handful of role files as a starting skeleton, rewrite the deliverables section to match your own definition of done, and ignore the rest of the catalog. The repository's structure supports that, because each role is an independent file with no cross-references that would break if you delete its neighbours.

## Maintenance, licence and what an upgrade actually costs

The repository is not archived, and the last push was on 2026-09-09. The most recent release listed is v1.2.6 from 2026-06-16, while package.json declares version 1.4.0, so the release tags and the package version are not in step. Treat that as a signal about how carefully to track versions: pin the version you install rather than following latest. Upgrades are cheap in the mechanical sense, because the artifact is text. They are not free in practice, because if you have edited role files in place, a reinstall or a fresh clone will not merge your changes. The safer pattern is to copy the roles you use into your own repository and treat agency-agents-zh as an upstream source you pull from deliberately. Licence terms are MIT, which permits commercial use and modification; the LICENSE file is at the repository root. That is a statement about the licence text, not legal advice, and the sponsor and course links in the README are commercial arrangements that do not change the licence of the role files themselves.

## Conclusion

Adopt it if you already drive Claude Code, Cursor or a similar tool and want a pre-written persona for a recurring job, especially if that job is Xiaohongshu, Douyin, WeChat or Feishu operations, where the 64 China-market roles have no obvious equivalent in the upstream English catalog. Do not adopt it if you expect a runtime: the repository is a content library, and the orchestration story lives in a separate project, agency-orchestrator, which the README links rather than bundles. Before committing, open the two or three role files you actually intend to use in the department directories, check that the deliverables match how your team writes them, and run npm run check:counts after any edit you make to a role count, since the repository ships a script whose only job is to catch that mismatch.

## FAQ

### How do I use agency-agents-zh in my project?

Install the npm package or clone the repository, then copy the role file you want into the instruction file your AI tool reads, such as Claude Code's project memory file or Cursor's rules directory. The repository does not mandate a single destination, and the examples directory shows roles being combined in workflows.

### What are Claude Code Agency agents?

In this repository they are Markdown role definitions that you place where Claude Code reads project instructions. Each agent carries its own persona, process and deliverables, and the README states the collection supports Claude Code alongside Cursor, Copilot and 18 other tools.

### What are the top 3 AI agents in agency-agents-zh?

The repository does not rank the 277 roles, and no ordering or quality tier appears in the README or the package metadata. The catalog files, CATALOG.md and AGENT-LIST.md, list what exists without ranking it.

## Sources

- [jnMetaCode/agency-agents-zh on GitHub](https://github.com/jnMetaCode/agency-agents-zh)
- [License: MIT](https://github.com/jnMetaCode/agency-agents-zh/blob/main/LICENSE)
- [Project website](https://ao.aiolaola.com/experts)
- [README](https://github.com/jnMetaCode/agency-agents-zh/blob/main/README.md)
- [Releases](https://github.com/jnMetaCode/agency-agents-zh/releases)

---

Hysen Labs editorial analysis, written from the project's own repository and release notes. Cite the canonical page: https://hysenlabs.com/projects/jnmetacode-agency-agents-zh