Preface

DSH’s philosophy is that everything is a plugin. When installing a third-party skill, there are two problems: making it discoverable by the harness, while not touching existing skill configurations—patching replaces the target configuration line entirely rather than merging it, and if you overwrite directly on the base provider line, the originally configured skill root gets wiped out.

riffkit/dsh-plugin handles exactly these two things: it packages Riffkit’s agent skill into a DSH bundle that is installable via a single command, registers its skill root using the insert method, and doesn’t touch existing configurations. Below is an introduction to its positioning, implementation, and usage.

What is this

riffkit/dsh-plugin, GitHub repository riffkit/dsh-plugin, package name @riffkit/dsh-plugin, current version 1.4.0, MIT license. Its positioning is written very conservatively—“A packaging layer, nothing more”: this package installs only one thing, Riffkit’s agent skill, and registers the package directory as an additional skill root via cordis.patch.yml. The skill source files are provided by riffkit.ai.

Riff stands for: analyzing why a winning short video is effective—hook, rhythm, emotional beats—and then rebuilding this recipe into new material using your product, persona, and language.

Core Capability: Inheriting Structure, Not Frames

Riff’s process consists of two steps: first analysis, then reconstruction.

  1. Analyze why the source video is effective: how the hook is established, how rhythm progresses, how emotional beats are laid out.
  2. Rebuild this recipe into new material using your product, persona, and language.

There is a boundary worth mentioning separately: the source video is never re-uploaded; it is the structure that is inherited, not any frame. What is in the new material depends on the elements you provide; the original footage will not be carried over.

Installation

In a DSH environment, installation is completed with a single command:

dsh plugin add @riffkit/dsh-plugin

After installation, manual registration of the skill root is not required; the registration logic is handled by the patch within the package, as covered in the next section.

Patch Design: Registering an Independent Root with Insert

The package.json of the package contains the dsh.bundle.patch field, pointing to ./cordis.patch.yml. The patch content registers a new skill root using the insert method:

# Patch content registers a new skill root using the insert method:
- insert:
    - id: riffkit-skill-root
      name: '@deepseek-ai/dsh-skill-filesystem'
      config:
        providerName: riffkit-bundle
        includeDefaultRoots: false
        customSkillDirs: [ …the skills/ directory inside the package… ]

Three key points:

  1. The provider uses DSH’s own @deepseek-ai/dsh-skill-filesystem, and the providerName is named riffkit-bundle;
  2. includeDefaultRoots: false, customSkillDirs only points to the skills/ directory inside the package, therefore this provider can only see this single root of its own;
  3. Using insert instead of overwriting the base provider line—the patch replaces the entire configuration of the target line rather than merging it in, and overwriting would clear the existing configurations you had there.

With this design, this provider does not obscure your existing skills: it cannot see your skills, and your skills will not block it.

Environments Without DSH

The README is straightforward: if you don’t use DSH, you don’t need this package. The skill itself is a Markdown file, and any harness that can read the skill root can load it directly:

mkdir -p ~/.agents/skills/riffkit
curl -sSL https://riffkit.ai/SKILL.md -o ~/.agents/skills/riffkit/SKILL.md

~/.agents/skills is the skill root scanned by multiple harnesses; placing a copy here allows all these harnesses to register with Riffkit. For the skill itself, examples, and installation commands for Claude Code, Cursor, and Codex, see github.com/riffkit/skill.

Typical Usage

Usage starts with a source: a TikTok link, an uploaded video, or an already analyzed template. The backend riffes the emotional recipe of this source into your own AI video—either a short video ready for direct posting or UGC-style ad material. Digital persona, product placement, and language are optional, provided as needed.

How Content Stays Current

There is no manually maintained content in this package; updates rely on a daily workflow:

  1. Pull SKILL.md, HEARTBEAT.md, and SKILL.json from riffkit.ai; this is the single source of truth, updated with application deployment;
  2. Derive package.json and cordis.patch.yml from these three files using scripts/generate.mjs;
  3. Verify that the derived results can still be discovered by DSH’s own provider using scripts/smoke.mjs;
  4. Publish versions missing in the registry.

Additionally, the skill comes with a 24-hour version heartbeat; installed copies will automatically resync between package updates.

Suitable Scenarios and Considerations

Suitable for:

  1. People working in DSH who want to turn a winning short video into their own material—provide a source, let the agent complete analysis and reconstruction;
  2. Those not using DSH; simply install the skill itself, no need to install this package.

Pre-installation considerations:

  1. The plugin runs with the permissions of the current dsh process; review the source code before deciding to install;
  2. The license is MIT, consistent with the license field in package.json at the end of the README;
  3. This package is only a packaging layer, and all capabilities come from the Riffkit skill itself; you can evaluate functionality directly at github.com/riffkit/skill.

Conclusion

Recap: riffkit/dsh-plugin does not do much, but does so with standard practices—a Markdown skill, one install command, an insert patch to register an independent root without overwriting existing configurations, a daily workflow plus a 24-hour heartbeat to ensure content doesn’t expire. If you want to riff short videos in DSH, or want to reference how to integrate an external skill into DSH, you can start here.

  • GitHub repository: https://github.com/riffkit/dsh-plugin
  • Community directory listing page: https://www.skillhub.cn/plugins/riffkit/dsh-plugin

The community directory is an independent third-party site with no official affiliation to DeepSeek.