dsh-h3-seg-prompt-design

Readme

# dsh-h3-seg-prompt-design

An installable [DeepSeek Harness](https://github.com/deepseek-ai/deepseek-harness) plugin bundle that registers the **`h3-seg-prompt-design`** skill. The skill turns a long-form H3 video idea into a locally organized, repeatable package of H3-compliant per-segment prompts.

H3 generates at most 15 seconds per call, so a long video is produced by looping many generation segments. This skill manages the whole loop: brainstorming, locking reference assets, designing 5–15 second segments, and writing the prompts for each segment.

## What it does

1. **Inspiration** — when you have no concrete concept, propose several distinct, segment-friendly directions.
2. **Project folder** — once a theme exists, create a folder named after the video theme inside the workspace.
3. **Asset mapping** — confirm your reference images / videos / audio, assign unique aliases, and write `素材映射.md` (paths may stay empty so prompts can be planned before assets exist; a reference video's picture and its companion audio are two independent assets).
4. **Storyboard** — expand the idea into 5–15 second segments, estimate each duration, mark spans that continue across segments with `（接续分镜NN）`, and write `分镜设计.md`. Enforces H3's per-generation caps: ≤9 images, ≤3 videos, ≤3 audio tracks.
5. **Prompts** — write `H3分镜提示词.md` following the official H3 guide, except reference assets are cited as `{{ref:别名}}` (companion audio as `{{ref:别名.audio}}`) instead of the official `<Picture N>` / `<Video N>` / `<Audio N>` numbering.

## Install

The bundle is a plain package with a `dsh.bundle.patch` declaration and no build step. Install it into a profile with `dsh plugin` (which forwards to pnpm):

```powershell
# From a local checkout (the plugin folder itself):
dsh plugin --profile web add .

# Or from a published git repo:
dsh plugin --profile web add github:<owner>/dsh-h3-seg-prompt-design
```

> Git-hosted plugins that ship a `prepare` build need pnpm's `allowBuilds` allowance; this bundle ships committed plain-JS, so no build allowance is required.

After the profile starts, the skill appears in the model's skill catalog. Use it by asking for segmented H3 prompts, or invoke it directly with `/h3-seg-prompt-design`.

## Layout

```text
dsh-h3-seg-prompt-design/
├─ package.json            # dsh.bundle.patch → ./cordis.patch.yml
├─ cordis.patch.yml        # mounts the provider plugin
├─ index.js                # Cordis plugin: registers the skill on ctx.skills
└─ skill/                  # the bundled skill (resourceBase)
   ├─ SKILL.md             # workflow + invariants
   ├─ references/          # workflow rules + the two official H3 guides
   ├─ assets/              # 素材映射 / 分镜设计 / H3分镜提示词 templates
   └─ scripts/
      └─ validate_project.py
```

## Notes

- The skill name is `h3-seg-prompt-design`; the npm package is `dsh-h3-seg-prompt-design`.
- `index.js` has **no runtime dependencies** — `skills` resolves as a Cordis service, so the package needs nothing beyond the dsh installation.
- Keep the provider `DESCRIPTION` in `index.js` in sync with the skill summary you want shown in the catalog.
- Replace `MIT` with your preferred license before publishing; no `LICENSE` file ships here.