rnskill: A 57-Skill Collection for Codex and Claude Code, Split Between Chinese Content Production and Coding

What rnskill is, and the specific gap it fills

rnskill is a collection of AI agent skills maintained by Pluviobyte, distributed as directories that each contain a SKILL.md file. The README describes it as a full set of skills for Codex, Claude Code and other agents that support project-level skills, and states there are currently 57 of them. The README is written in Chinese with an English translation at README.en.md, and the skill names are Chinese, which tells you who the intended reader is.

The problem it addresses is not model capability. It is that a video production workflow has many discrete steps, and each step needs its own instructions, its own input contract and its own quality gate. Downloading a Douyin clip, extracting a word-level transcript, rewriting the script, generating a cloned voice, driving a digital avatar, burning subtitles and producing a cover are seven different jobs. rnskill turns each into a skill the agent loads on demand rather than one long prompt. The README's own example of the payoff is a production claim: under 10 hours of total input over one month, 2k new followers on Douyin and a first paid brand deal. That is a self-reported outcome from the maintainer, not a benchmark, and it depends on the operator as much as the skills.

Who it is for: someone running an agent-driven Chinese short-video or article pipeline who is comfortable installing skills from a third party and wiring up external tools. Who it is not for: an English-only team, or anyone who wants a curated coding-skill pack today.

The SKILL.md contract and how skills get discovered

The mechanism is deliberately plain. A skill is a directory under skills/<skill-name>/ containing a SKILL.md, and the agent reads it from a project-level path. The README states the prerequisite as: Codex, Claude Code or another agent that supports project-level skills, and the target project must be able to read .agents/skills/<skill-name>/SKILL.md. That path is the Codex convention. Claude Code uses .claude/skills/<skill-name>/ instead, which is why the manual install section gives two different copy commands.

There is no runtime, no daemon and no plugin API beyond the agent's own skill loader. The composition happens two ways. The first is explicit invocation by name, as in the README's example: use $rn-cover-skill to make a 5:2 cover with the title given. The second is chaining declared inside the skill text. ra-洗稿 is described as automatically chaining ra-人话, then dbs-ai-check, then dbs-hook, then dbs-resonate, then ra-video-title. ra-video-wash-pipeline is a dispatcher that strings download, transcript extraction, rewriting, quality check and queueing. ra-video-production-director reads a handoff contract, calls downstream skills, and manages state, archiving and QC.

That is a shallow architecture, and the trade-off is visible. Because every skill is just a markdown file plus whatever scripts sit beside it, there is no shared schema enforced at runtime. The handoff contract mentioned for the production director is a convention the skills agree on, not something the agent validates. If you edit one skill in a chain, nothing tells you the next one still matches.

Installing it: marketplace, npx, or a plain copy

Three install paths appear in the README. The Claude Code plugin marketplace route is two commands:

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

The generic route uses the skills CLI and installs everything globally:

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

To take a single skill instead of all 57:

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

Manual installation is a copy into the agent's project directory. For Codex:

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

For Claude Code, the same idea with a different target:

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

The --all flag is the part worth pausing on. Fifty-seven skills installed globally means 57 sets of instructions available to every session, and some of them carry hard rules. tts-skill is described as forbidding cloud fallback and requiring a fixed lossless reference plus a voice_manifest.json. heygen-digital-avatar mentions a hard gate on audio confirmation. Installing the whole set to try one cover generator puts those constraints in front of the agent whether or not you wanted them. The single-skill install exists precisely for that reason, and it is the one I would use first.

The video pipeline skills, and where the external dependencies bite

The content-creation half is the substantial half. Reading the table, the chain runs roughly like this. ra-video-download pulls from Douyin, YouTube, Bilibili, Twitter and Xiaohongshu, with yt-dlp and TikHub named as the underlying tools. ra-逐字稿提取skill takes Douyin and Xiaohongshu video and produces a transcript using watermark removal plus Paraformer ASR. ra-洗稿 rewrites the script and chains the quality skills. tts-skill does local voice cloning with IndexTTS2. heygen-digital-avatar handles Digital Twin generation and compositing through HeyGen's CLI OAuth. ra-audio-to-subtitles uses Volcengine Doubao-ASR for word-level timestamps and runs alignment, fragmentation, connector and reading-speed checks. skill-captions renders and burns subtitles in anchor-dark or anchor-light with 4K native redraw and render QC.

The dependency list is long and mostly not bundled. yt-dlp, TikHub, Paraformer, IndexTTS2, HeyGen credentials, Doubao-ASR and HyperFrames are all named as the machinery underneath. The README does not document how to install or authenticate most of them. heygen-digital-avatar is the exception in that it names CLI OAuth as its auth path, and tts-skill is the exception in that it names a manifest file. If you do not already have IndexTTS2 running locally, the voice-cloning skill is a description of a workflow you cannot execute.

The language coverage is also a constraint rather than a detail. Paraformer ASR and Doubao-ASR are Chinese-oriented services, the skill names are Chinese, and several skills exist specifically to strip AI-flavoured Chinese prose. ra-人话 is described as removing binary-contrast shells, pseudo-insight markers and colon-lecture tone while preserving the author's judgement and facts. That is a Chinese writing problem. Pointing it at English copy is not what it was built for.

The dbs toolbox and the licence that is not settled

Twenty-two of the 57 skills come from dbskill, an open-source project by @dontbesilent, and the README states they are under CC BY-NC 4.0. That is a non-commercial licence. It covers dbs itself (routing before and after a task), dbs-hook for video-opening diagnosis paired with ra-hook, dbs-resonate for resonance diagnosis, dbs-ai-check for AI writing traces, dbs-content, dbs-spread, dbs-diagnosis, dbs-benchmark, dbs-goal, dbs-deconstruct and dbs-action.

Separately, the repository metadata reports the licence as NOASSERTION. That is the GitHub label for a licence it could not identify automatically, and it means the top-level terms of the repository are not stated in a machine-readable form. The README does give per-skill provenance: the ⬡ marker means a skill comes from or is adapted from an external open-source project, and the table lists MIT for video-use and ian-xiaohei-illustrations, Apache-2.0 for chengfeng-videocut-skills, AGPL-3.0 for ai-jian-koubo, and CC BY-NC 4.0 for the dbs set. editorial-collage-motion is described as a locally free compatible version inspired by an external collage-motion skill.

So the picture is a mixed-licence bundle without a single top-level statement. I am not a lawyer and this is not legal advice, but the practical consequence is concrete: if you plan to use any of this commercially, the 22 dbs skills are the ones to check first, because CC BY-NC 4.0 restricts commercial use regardless of what the rest of the repository says. The AGPL-3.0 skill, ai-jian-koubo, is a rewritten and extended version of chengfeng/videocut-skills, and AGPL carries its own obligations if you distribute a modified version. Opening the LICENSE file and the per-skill directories before adoption is the only way to know what you are agreeing to.

rn-cover-skill as a worked example of the design

The README singles out rn-cover-skill, and it is the clearest illustration of how these skills think. It generates a 5:2 editorial-diagram cover from a title and topic with no reference image required. The design splits the work into two layers: ImageGen redesigns the right-hand concept graphic each time, while a script handles exact typography, the warm-white background and an editable SVG. The stated reason is that this keeps a family resemblance across covers without copying one reference cover's nodes and composition over and over.

Specifics from the README: a fixed #FAF9F5 warm-white canvas, Chinese or mixed titles preferred, left-side text vertically centred as a block, the right-side graphic yielding space based on title length, a fine grid on the right that stays visible at thumbnail size, and adjustable origin, spacing, intensity and graphic position. Output is a self-contained SVG plus a PNG. Invocation is one line: use $rn-cover-skill to make a 5:2 cover, title given.

Two other cover skills sit alongside it. skill-cover registers styles, produces dual-ratio assets and generates images automatically. editorial-dot-cover uses a warm-grey paper base, oversized Chinese title, whitespace and dot-matrix vector icons, outputting SVG and PNG. The overlap is real; the difference is that rn-cover-skill is the one documented as not needing a reference image. If you are choosing between them, that is the deciding line.

Limitations, failure modes and the wrong-tool case

The most obvious limitation is that the coding category is thin. The README has a section headed coding development describing it as general coding skills, continuously being added. That is the whole entry. Of the 57 skills, the ones I can trace to coding are essentially the dbs business-diagnosis set and grok-build-cli, which lets Codex call a local Grok Build CLI, check login and model, choose single-turn or agentic calls, and poll long jobs. If you came for a coding-skill library, you are early.

The second limitation is external service dependence. HeyGen, TikHub, Volcengine Doubao-ASR and HyperFrames are named as part of the pipeline. Any of them changing its API, pricing or terms breaks the corresponding skill, and the repository cannot fix that for you. The local alternatives are local only in part: IndexTTS2 runs locally, Paraformer runs locally, but the download and subtitle paths do not.

The third is quality-gate subjectivity. rn-replica-qc describes five fidelity levels and three full-frame gates across assets, runtime and delivery. rn-motion-director mentions an Anti-PPT quality gate. These are judgement calls encoded as instructions, and the README does not give pass thresholds. Two operators running the same gate can reach different verdicts.

The wrong-tool case is a team that wants a deterministic render pipeline. If you need byte-identical output from the same input, a set of markdown instructions driving an LLM is the wrong layer. Use it to plan and orchestrate; put the deterministic rendering in a script.

How it compares with plain skills and with the upstream projects

The nearest alternative is not a competing product, it is writing your own SKILL.md files. A single skill for your own video workflow is maybe a page of instructions, and the agent needs no third-party install. The difference in approach is scope versus fit. rnskill gives you 57 directories covering选题, hook selection, script rewriting, TTS, avatar, subtitles, covers, motion and review, with the chaining already declared between them. Writing your own gives you exactly your steps and nothing else, at the cost of building each one. The rnskill route makes sense when your workflow resembles the one described; it makes less sense when it does not.

The other comparison is with the upstream projects the README credits. chengfeng-videocut-skills is included at its original version under Apache-2.0: transcript, error detection, review page, FCPXML export. ai-jian-koubo is a rewrite and extension of the same idea under AGPL-3.0, adding AI error detection and a web waveform review step. If you only need the transcript-to-FCPXML path, taking the Apache-2.0 original avoids the AGPL obligations of the rewrite. For the dbs set, the upstream is dbskill by @dontbesilent; rnskill is a redistribution, so the licence and the maintenance both live upstream. For ian-xiaohei-illustrations, the README credits Ian / helloianneo and MIT, and describes the skill as turning article judgements, flows and metaphors into 16:9 illustrations in a white hand-drawn style with sparse red, orange and blue annotation. Going to the original gets you the same content with a cleaner provenance trail.

Maintenance cost and upgrade surface

The repository is not archived, the default branch is main, and the last push recorded in the metadata is 2026-08-30. The most recent release is v1.1.0 from 2026-07-14. That is roughly six weeks between the release and the last push, which suggests active work between releases rather than a frozen snapshot. The README describes the coding category as continuously being added, so skill count and contents should be expected to move.

For an adopter, the upgrade surface is the whole set. Installing with --all and later re-running the same command will pull whatever the current state of the repository is, including new skills and edits to ones you already rely on. Because the chains are declared in prose inside each SKILL.md, an upstream edit to ra-人话 can change the behaviour of ra-洗稿 without any version bump on the latter. There is no lockfile and no per-skill versioning in the material I have. If you pin a working set, the practical move is to copy the skill directories into your project as the manual install section shows, and treat the copy as your version rather than tracking main.

The release history is short. One release is listed, v1.1.0, so there is no track record of how breaking changes get communicated. That is a fact about the material, not a criticism of the project's intentions.

Editorial conclusion

Adopt rnskill if you already run Codex or Claude Code against Chinese-language short-video production and want the pipeline steps (download, transcript, rewrite, TTS, avatar, subtitles, covers) expressed as separate SKILL.md directories you can install one at a time. Do not adopt it as a general coding-skill library: the README lists the coding category as still being filled in, and the bulk of the 57 skills assume Douyin, Xiaohongshu, IndexTTS2, HeyGen or HyperFrames. Before you install anything, open the LICENSE file, because the repository metadata reports NOASSERTION while the README states that the 22 dbs skills come from dbskill under CC BY-NC 4.0, a non-commercial licence that will not suit every user.

Official sources

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