Perform evidence-driven verification of candidate resources. Do not make decisions based only on repository descriptions, search snippets, or Stars.
Workflow
1. Normalize the candidate address
- Accept a GitHub repository, repository subdirectory, PR, Issue, or a project name provided by the user.
- From
tree/、blob/and other subpaths, extractowner/repo, while preserving the original subpath that points to an independent Skill. - Search the current curated list for the repository name, normalized URL, and former names to determine whether it is a new addition, duplicate, or update to an existing entry.
- When a repository is renamed or transferred, use the final canonical address to avoid continuing to include redirect URLs.
rg -n "owner/repo|项目名" README.md
2. Obtain repository facts
Prefer using an authenticated GitHub CLI to obtain read-only information:
gh repo view owner/repo \
--json nameWithOwner,description,stargazerCount,forkCount,isArchived,isPrivate,createdAt,pushedAt,licenseInfo,url
gh api repos/owner/repo/contents \
--jq '.[] | [.type,.name,.path] | @tsv'
gh api repos/owner/repo/contents/README.md \
-H 'Accept: application/vnd.github.raw+json'
If gh is unavailable, public web pages or other read-only methods may be used for verification, but you must state which facts could not be confirmed.
3. Verify real assets and installation paths
Confirm that at least one reusable asset actually exists:
SKILL.mdorskills/.claude-plugin/plugin.json, marketplace or plugin installation commandagents/、commands/, hooks, or migratableAGENTS.md- MCP Server, CLI, SDK, or an explicit Agent workflow
- Installation instructions for tools such as Claude Code, Codex, Cursor, Gemini CLI, and OpenClaw
Continue checking:
- Whether the README explains prerequisites, installation, configuration, and usage.
- Whether example commands point to real package names or paths.
- Whether the repository is empty, contains only conceptual copy, or has only a marketing page.
- Whether a subdirectory Skill can be installed independently; if it cannot be installed independently, prioritize including the repository-level address.
4. Evaluate value for Chinese users
Judge value from actual scenarios, not just by the number of features:
- Whether it addresses high-frequency needs such as code review, testing, security, documentation, performance, DevOps, content creation, office automation, or knowledge management.
- Whether it supports Chinese content, domestic platforms, the Windows environment, or common workflows of Chinese teams.
- Whether an English project has a sufficiently clear installation path so that Chinese users can use it without additional research.
- Compared with already included projects, whether there are new scenarios, official sources, or a clearly more complete implementation.
Suggested categories can only be selected from the project's existing categories:
- Star skills
- Platform operations / self-media / office traffic
- Development efficiency
- Content creation
- Academic research
- AI Agent
- Finance / business
- Chinese-exclusive
5. Check risks and mandatory elimination conditions
In the following cases, directly recommend not including:
- 404, private, empty repository, obvious template, or archived with no replacement address.
- No direct relationship with Agent Skills, Plugin, MCP, or AI programming workflows.
- No visible actual assets, installation method, or verifiable capabilities.
- Involves account theft, malicious attacks, bypassing payment, illegal scraping, or obvious infringement.
- The project description clearly does not match the repository contents.
The following cases are not necessarily eliminated, but must be explicitly flagged:
- Strong dependence on commercial APIs, subscriptions, cloud browsers, or third-party accounts.
- Automatically publishes content, sends messages, executes transactions, or modifies the state of external platforms.
- Involves crawlers, login sessions, Cookies, Tokens, browser automation, or account risk control.
- License is unclear, or restricts commercial use, redistribution, and modification.
- Just created, low Star, long-term unmaintained, or supported only by a single unverified service.
6. Score and give a conclusion
Use a 100-point scale to aid judgment; evidence and hard requirements take precedence over the total score:
| Dimension | Score | Judgment focus |
|---|---|---|
| Directly related to the Agent ecosystem | 25 | Explicitly supported tools and workflows |
| Installability | 20 | Install commands, prerequisites, reproducibility |
| Real assets | 20 | Skill, Plugin, Agent, MCP, CLI, etc. |
| Documentation completeness | 15 | README, examples, configuration, and boundaries |
| Value for Chinese users | 10 | High-frequency scenarios, Chinese or domestic ecosystem adaptation |
| Activity and credibility | 10 | Updates, maintainers, license, tests |
Deduct 0–30 points based on risk. Default interpretation ranges:
75–100: recommended for inclusion60–74: include with caution, boundaries or evidence need to be supplemented<60: not included for now
Do not automatically reject due to low Stars, and do not skip asset verification due to high Stars.
Output format
## 候选资源核验报告
### 结论
- 建议:收录 / 谨慎收录 / 暂不收录 / 更新现有条目
- 推荐分类:分类名
- 总分:XX/100
### 核验事实
| 项目 | 证据 |
|------|------|
| 仓库状态 | 可访问、未归档、最近更新时间 |
| 实际资产 | 具体路径或文件 |
| 安装方式 | 已验证的命令或文档位置 |
| 支持工具 | Claude Code / Codex / ... |
| 重复检查 | README 中是否已有 |
### 风险与边界
- 前置条件、账号风险、商业依赖或许可证限制
### 中文收录文案
| [项目名](https://github.com/owner/repo) | 解决什么场景,支持哪些工具,对中文用户有什么价值(123⭐)|
### 核验说明
- 未确认的事实和需要维护者人工复核的部分
Inclusion execution rules
- When the user only requests an evaluation, output only the report and do not modify the repository or GitHub status.
- When the user explicitly requests adding to the curated list and the candidate passes verification, first modify the README, then run the official website sync and project checks.
- Use natural Chinese for descriptions, keep technical terms in English, and do not copy the English repository introduction verbatim.
- Use the real Star value at the time of verification; if it cannot be read, omit it and explain, and do not guess.
- Only use words such as "official", "production-grade", "largest", and "first" when there is sufficient evidence.
- The inclusion copy for high-risk projects must clearly state dependencies and applicable boundaries, and must not imply risk-free fully automated operation.