rnskill: a 57-skill library for Codex and Claude Code agent workflows

What rnskill is for, and who it is not for

rnskill is a collection of AI Agent Skills maintained by 雪踏乌云. The README states there are 57 skills, split into content creation (video, image-text, articles) and coding development. A skill here is a directory containing a SKILL.md file that an agent reads from a project-level path such as .agents/skills/<skill-name>/SKILL.md. The README lists Codex and Claude Code as the supported agents.

The audience is narrow and specific. The content-creation half assumes a Chinese-language workflow: 抖音 and 小红书 as distribution targets, 微信公众号 extraction, Chinese copy de-AI-ification (ra-人话), Chinese TTS voice cloning through IndexTTS2, and HeyGen digital avatars. If your content pipeline is English-language, most of these skills solve problems you do not have. The coding half is described only as "通用编码 Skill，持续补充中", which is the vaguest line in the document and gives no detail on what those skills do.

The README also makes a personal claim about the maintainer's own use: under 10 hours of total effort in a month, 2k followers gained on 抖音, and a first paid commission. That is a self-reported anecdote, not a benchmark, and it does not describe what any other user should expect.

How a skill actually reaches the agent

There is no runtime, daemon or server. The mechanism is file placement plus agent-side interpretation. The README states the prerequisite plainly: the target project must be able to read .agents/skills/<skill-name>/SKILL.md. The agent discovers the skill by reading that file, and the file's instructions then drive the agent.

That design has a consequence worth naming. Because skills are prose instructions plus whatever scripts ship alongside them, the quality ceiling is set by the SKILL.md text and by the external tools each skill calls. The README shows how deep those dependencies go. ra-video-download wraps yt-dlp and TikHub. ra-逐字稿提取skill uses Paraformer ASR. tts-skill uses local IndexTTS2 and the README says it forbids cloud fallback and leaves a voice_manifest.json. heygen-digital-avatar uses HeyGen's CLI with OAuth. ra-audio-to-subtitles uses 火山 Doubao-ASR for word-level timestamps. None of these are bundled; the agent is expected to reach them on the machine.

The repository layout is flat and readable: .claude-plugin/, .github/, skills/, tools/, docs/, assets/, plus CREDITS.md, LICENSE, VERSION, README.md and README.en.md at the top level. The skills/ directory is the product. The .claude-plugin/ directory is what makes the Claude Code marketplace install work.

Installing rnskill and generating a first cover

The README gives three install routes. The Claude Code plugin marketplace route uses two commands, one to register the marketplace and one to install a single skill named ra-人话:

claude plugin marketplace add Pluviobyte/rnskill
claude plugin install ra-人话@rnskill

The generic route works for Codex and Claude Code and pulls everything globally with --all:

npx -y skills add Pluviobyte/rnskill -g --all

To install one skill instead of the whole set, the README shows the same command with a --skill flag:

npx -y skills add Pluviobyte/rnskill --skill ra-人话

Manual installation means copying a skill directory into the agent's expected path. The README gives both variants:

# Codex
mkdir -p <project>/.agents/skills
cp -R skills/ra-人话 <project>/.agents/skills/ra-人话

# Claude Code
mkdir -p <project>/.claude/skills
cp -R skills/ra-人话 <project>/.claude/skills/ra-人话

After that, the first real use in the README is the cover skill. You invoke it by name inside the agent conversation, giving it the ratio and the title:

使用 $rn-cover-skill 制作 5:2 封面，标题为「开源 Claude 风格封面 Skill」

What you should see, per the README: a 5:2 editorial illustration cover on a fixed #FAF9F5 warm-white canvas, with the title block vertically centred on the left and a freshly generated concept graphic on the right, delivered as a self-contained SVG plus a PNG. The README notes the right side keeps a faint grid visible in thumbnails, and that the grid's start point, spacing and intensity can be adjusted per cover. Note that this skill does not require a reference image; the README says the ImageGen layer redesigns the right-hand graphic each time while the script handles typesetting and the editable SVG.

Licence status is the real adoption risk

The repository's licence field is NOASSERTION, which means no standard licence was detected. That is not the same as public domain, and it is not the same as unlicensed. The README's own attribution column shows why the field is unresolved: the collection is a mix. Original skills are labelled 原创. Others are marked with ⬡ and carry their upstream terms: video-use is adapted from Browser Use under MIT, chengfeng-videocut-skills is Apache-2.0, ai-jian-koubo is AGPL-3.0, ian-xiaohei-illustrations is MIT, and the 22 dbs business skills come from @dontbesilent's dbskill project under CC BY-NC 4.0.

CC BY-NC 4.0 is a non-commercial licence. If you are evaluating rnskill for commercial content production, the dbs subset is the part to examine first, and the README's own table is the place to do it. AGPL-3.0 in ai-jian-koubo carries obligations that differ from MIT. This is not legal advice; it is a description of what the repository's own files say, and the practical step is to read LICENSE and CREDITS.md before shipping anything built on these skills.

Where rnskill breaks down

The dependency chain is the first failure mode. Skills such as tts-skill, ra-逐字稿提取skill and ra-audio-to-subtitles assume specific external services and models on the machine. The README does not document what happens when IndexTTS2 is absent, when the HeyGen OAuth token expires, or when a TikHub call fails. For tts-skill the README states cloud fallback is forbidden, which means no local model means no output rather than a degraded one. That is a deliberate choice, and it also means the skill is unusable on a machine without the local model installed.

Language lock-in is the second. The skill names, the invocation examples and the writing skills are Chinese-first. ra-人话 targets 中文去 AI 味 writing and hard-bans specific Chinese rhetorical patterns. Running it on English copy is not what it was built for.

The third is documentation depth. The README is a catalogue with one-line descriptions, not a manual. For most of the 57 skills there is no documented input format, no error behaviour and no rollback path. The README does not document rollback for any skill. If you need to know what a skill does to your files before it does it, the README will not tell you; you have to read the SKILL.md.

Finally, treat the whole set as a moving target. The last push to the repository was on 2026-08-30, and the most recent release listed is v1.1.0 from 2026-07-14. A large skill collection that changes frequently is a collection whose individual skill behaviour can change between installs.

rnskill versus a single-purpose skill repository

The obvious alternative is not a different product but a different shape: install one skill from a focused repository rather than a 57-skill bundle. rnskill itself supports this through the --skill flag and through manual copying of a single directory, so the comparison is really about whether you want the catalogue or the item.

A focused repository gives you one SKILL.md, one licence, and one dependency list to audit. rnskill gives you breadth across a production pipeline that is designed to hand off between stages: ra-选题 routes to a workbench, ra-video-wash-pipeline chains download, transcription, rewriting, quality check and queueing, and ra-video-production-director reads a handoff contract and calls downstream skills. That chaining is the actual value proposition, and it only works if you install the connected pieces. Installing rn-cover-skill alone gets you a cover generator; installing the pipeline gets you a production line.

The honest framing is that these are two different bets. A single-skill repo is easier to audit and easier to abandon. rnskill's pipeline skills are worth the larger surface only if you intend to use the handoffs.

Questions worth answering before you install

Start with the licence question, because it gates everything else. Read LICENSE and CREDITS.md, then check whether the specific skills you need fall under 原创, MIT, Apache-2.0, AGPL-3.0 or CC BY-NC 4.0. The dbs subset is the one to look at closely for commercial use.

Next, check the dependency reality. If your work needs tts-skill, confirm IndexTTS2 is installed locally, because the README says cloud fallback is forbidden. If your work needs heygen-digital-avatar, confirm the HeyGen CLI OAuth flow works in your environment. If your work needs ra-video-download, confirm yt-dlp and TikHub access. These are the points where an install succeeds but a run fails.

Then decide on install scope. The --all flag is convenient and pulls all 57 skills into your global agent configuration; the --skill flag and manual copy keep the surface small. For a first evaluation, a single skill copied into .agents/skills or .claude/skills tells you more about whether the SKILL.md format suits your workflow than a full install does.

Finally, note the version marker. The repository carries a VERSION file and the latest listed release is v1.1.0 from 2026-07-14, while the last push was 2026-08-30. Pin to what you have read rather than assuming the skills you install today match what you read tomorrow.

Editorial conclusion

Adopt rnskill if you already run Codex or Claude Code and your work is Chinese-language short video, article or cover production, because the skills assume that pipeline and its tooling. Do not adopt it if you need a stable, versioned dependency: the licence is NOASSERTION, and the README's own attribution table shows the collection mixes original skills with adapted third-party code under MIT, Apache-2.0, AGPL-3.0 and CC BY-NC 4.0 terms. Before installing, open skills/rn-cover-skill/SKILL.md and one dbs skill to confirm the licence and dependency expectations match your project, then decide whether the whole set or a single --skill install is what you want.

Frequently asked questions

Official sources

1. Issues
2. Pluviobyte/rnskill on GitHub
3. README
4. Releases