redfox-community: Agent Skills for Chinese Social Media Research

The gap redfox-community fills: agent skills that call live Chinese platform data

Most Agent Skills collections are prompt libraries. They give an agent a procedure and a tone, and the agent still has no way to see what is actually trending on Douyin this morning. redfox-community takes a different route. Each skill folder bundles a SKILL.md entry point with optional scripts and reference documents, and the skills are written for scenarios the README names explicitly: inspiration, topic selection, copywriting, and data review. The topics list adds the platforms: douyin, xiaohongshu, redfox, social-media. So the intended user is someone running an agent over Chinese short-video and social platforms, who wants the agent to fetch real posts and account data rather than improvise. A second audience exists: developers who want to see how a vendor structures Agent Skills for a commercial API, since this repository doubles as a reference layout. The README is bilingual, with README.en.md linked from the top of README.md, which suggests the maintainers expect non-Chinese readers to at least evaluate the structure.

Folder-per-skill and SKILL.md frontmatter: the actual contract

The repository layout is deliberately flat. Under skills/ each skill gets its own subdirectory, and that subdirectory must contain SKILL.md. Everything else is optional and lives beside it: references/, scripts/, assets/. The README states the goal directly, that a single skill should be self-contained and use relative paths. That constraint matters more than it looks. If a skill references a file through an absolute path or reaches into a sibling skill's directory, copying the folder into Cursor's user skills directory or a project's .cursor/skills/ path breaks it. The frontmatter contract is minimal: name and description. The README's example uses a one-line description and advises it be specific about the scenario and the capability boundary rather than vague. There is no schema for inputs, outputs, required API endpoints or credential scope. That is a real weakness. An agent reading only the frontmatter cannot tell whether a skill needs a RedFox API key, which platform it queries, or what it returns. The body of SKILL.md has to carry all of that, and the README does not prescribe a body structure beyond mentioning trigger conditions, steps, constraints and examples.

The API key is the dependency the skills do not declare

The README has a short section titled 身份认证 (authentication) stating that all API requests require a valid API key, with a link to the RedFox hub settings page for issuing one. Nothing in the repository structure captures that. There is no manifest field for credentials, no environment variable name given in the material, and no indication of which skills need the key versus which are pure prompt procedures. In practice this means a skill can install cleanly, appear in the agent's skill list, and fail at the first API call with an auth error the agent may not interpret correctly. If you are evaluating this repository, treat the API key as a prerequisite you must provision before any skill that touches platform data will do anything. The README also points to a multi-platform API documentation set covering Douyin, Xiaohongshu, WeChat public accounts, WeChat Channels, Kuaishou, Bilibili, Toutiao and TikTok, with per-endpoint pages for things like account info, work lists, keyword search, comment retrieval and video-to-transcript submission. The skills are the thin layer over that catalogue.

Three install paths, and the one that actually resolves dependencies

The README documents four ways to get a skill in place. The first is manual: copy the skill subdirectory into your tool's skills directory, for example Cursor's user skills location or a project-level .cursor/skills/, or point the client's add-skill entry at the folder. The second is to ask an agent to do it. The README gives a prompt template that tells the agent to check whether the skill is already installed, read SKILL.md and the README at the GitHub URL, install into the project's existing skills path if absent, diff and update if present, then report the install path, the trigger and a short usage example. The third is the skills CLI, with three commands quoted in the README: npx skills init to install the CLI, npx skills add redfox-data/redfox-community to browse and pick a skill from the repository, and npx skills add followed by a full tree URL such as the seedance-video-gen skill to install one specific skill. The CLI then asks whether to install into a named agent folder or globally. The fourth is through third-party hubs: SkillHub, where you search by the skill's Chinese name, and ClawHub, where the README points to the redfox-data user page. The manual copy path is the one that quietly drops scripts and references if you copy only SKILL.md. The CLI path is the one that keeps a skill's folder intact.

Where this repository is thin, and when it is the wrong tool

The supplied material shows no LICENSE file, and the license field is listed as unknown. For a repository whose skills call a commercial API, that is the first thing to resolve before any internal adoption, because it determines whether you can fork a skill, vendor it into a private repo, or redistribute a modified version. The README does not address it. A second limitation is versioning. There are no releases in the supplied material, and the README's agent prompt template handles updates by diffing remote content against the local copy. That works for one skill and gets tedious across a set, because there is no version number in the frontmatter for the agent to compare. Third, the skills are platform-specific by design. If your work is on English-language platforms, or on any platform absent from the API list, the skills have nothing to query. Fourth, and most important: this is not an offline analysis toolkit. Every substantive skill depends on a live authenticated request to RedFox infrastructure, so if you need reproducible analysis over a frozen dataset, or you cannot send query terms to a third-party service, the architecture is the wrong fit regardless of how good the prompts are.

Compared with a general-purpose skills collection

A general collection such as Anthropic's own skills examples or a community prompt library ships skills that operate on files, shell commands and the model's own knowledge. They have no external service dependency, which is why they install and run with nothing but the agent. redfox-community inverts that trade. You get skills that can answer questions a prompt-only skill cannot, like what a specific Douyin account posted recently or which Xiaohongshu notes are drawing comments, because the data comes from RedFox's API endpoints rather than from the model's training cutoff. The cost is that every such skill inherits an account, a key, a rate limit you cannot see from the repository, and a vendor whose terms are separate from the repository's. The two are not substitutes. A file-manipulation skill and a Douyin account lookup skill solve different problems, and a team building a content workflow will likely run both. The honest framing is that redfox-community is an integration layer, and integration layers are judged by the service they integrate, not by the folder structure.

Maintenance, updates and what the licence question means for forks

The repository was last pushed in September 2026 according to the metadata, and the description promises continuous updates. No releases were retrieved, so there is no changelog to read and no tagged version to pin. That shapes the upgrade path: you either re-run the skills CLI add command against the repository, or you follow the README's agent prompt and let the agent diff remote content against your local copy. Both approaches work, neither gives you a stable version identifier, and neither tells you what changed between two pulls beyond a file diff. For a team, the practical consequence is that you should vendor skills into your own repository if you need reproducibility, and record the commit you vendored from. On licensing, the material does not state a license, so treat redistribution as unresolved until you check the repository for a LICENSE file yourself. This is a factual gap in the supplied information, not a legal opinion.

Editorial conclusion

Adopt redfox-community if your agents already work on Douyin, Xiaohongshu, WeChat or Bilibili content and you are willing to hold a RedFox API key. Skip it if you need offline analysis or a permissive licence you can verify. Before installing, read skills/<skill-name>/SKILL.md for the skill you actually want, confirm the frontmatter name and description, and check the repository for a LICENSE file, which the supplied material does not show.

Official sources

1. Issues
2. README
3. redfox-data/redfox-community on GitHub